The independent AI pricing desk

AI pricing, measured
for your workload.

Compare model costs. Track price changes. Free, no signup.

22 providers Checked nightly Standard input rates up to 250× apart per 1M tokens

5 OctNo price moves overnight · 16 of 22 providers verified, the rest carried forwardHistory →

The whole idea, in plain English

AI companies charge for the words you send and the words they send back. Text sent from files can count too. For image and video generation pricing, use PixelScale. Everything is measured in tokens (about ¾ of a word each). Every model prices tokens differently and suits different jobs: a quick model might cost a fraction of a cent, a sharper one fifty times more for the same request.

TokenScale is for the people deciding what to build with AI and what it will cost: curious first-time users, engineers choosing an API, product teams planning a feature, and researchers or leaders checking the market. Start with a familiar task, then compare the trade-offs in plain numbers.

Where these prices come from

Prices are checked nightly against provider sources. Any value that cannot be reproduced is carried forward and marked in the public record.

✓

How we verifyNightly checks, misses included

📈

How prices movedEvery model ranked, at every price change

🌱 Your 20-second tourThree small steps, with your choices kept visible all the way through.

Pick something you'd write

AI companies charge by the word, more or less. Let's see what real writing actually costs.

Technically they charge by the token, a chunk of a word, about three quarters of one. 100 words is roughly 133 tokens. That's the only jargon you need today.

3 /22

Claude Sonnet 5.5

Anthropic

$0.00186

under a penny

Play with AI pricing.Tap the wheel to lock in a model.

At 1,000,000 emails a day, that's $69.16 to $5,320 depending on the model.

or keep exploring below

spin · pick a task · tap centre

Try another provider and watch the chart change Try another provider

See how a conversation compounds 10 replies re-read the growing history

Live from your picks above

Every reply re-reads the entire conversation

200 words of context with 150-word replies. Change the wheel provider or size above and this redraws.

Anthropic Claude Sonnet 5.5 Email · 200 words

Current mid-tier rates: $2 input / $10 output per 1M tokens. Choose another provider in the wheel to compare.

$0.00293

$0.00332

$0.00372

$0.00412

$0.00452

$0.00492

$0.00532

$0.00572

$0.00612

R10

$0.00652

10 replies total$0.04722

Context re-read (grows)

Output written (fixed)

What does this cost at scale? Use the same selected workload at monthly volume

📈 Scale your selection

Take the exact 10-reply conversation above, then choose how many times it runs in a month.

10-reply conversations per month3,000

Your selected model

Anthropic · Claude Sonnet 5.5 $141.65 / month

3,000 email conversations, each with 10 replies

Across all 66 standard-rate model tiers, that same monthly workload ranges from $1.06 to $384.04 . Same volume, different model.

Explore the market DeepInfra leads the cheapest mid-tier list this week

📊 What leads this week Cheapest mid-tier model for a typical job · rank change vs 7 days ago

1DeepInfra$0.000878—
2Cohere$0.00140—
3Fireworks AI$0.00140—
4Groq$0.00140—
5OpenRouter$0.00233—

Compare all 22 providers →

Creator picks

Data determines the charts. Humans add colour. See the picks →

Common questions

What does it actually cost to run an AI API call?

It depends on the model and how much text goes in and out. TokenScale shows real examples. Writing the whole of The Hobbit costs about $0.32 on Gemini Flash-Lite 3.5 at list rates as of 5 Oct 2026, and about $0.0051 on the cheapest tracked tier. The same job costs far more on a frontier model. Pick a content size and a provider to see the live figure.

Why does the output price matter more than the input price?

Most providers charge more for the tokens a model generates than for the tokens you send. Output often costs three to five times more than input. A long answer to a short question can cost more than it looks. TokenScale splits every price into input and output so the gap is visible.

How much do AI prices differ between providers?

A lot. Across the 22 providers TokenScale tracks, input rates span roughly 250×, from $0.02 to $5 per million tokens as of 5 Oct 2026, and even models aimed at similar work are routinely more than thirty times apart. Choosing the right provider for a task can cut the bill by an order of magnitude. The comparison table sorts every provider cheapest first.

What is the cheapest way to run a large AI workload?

Pick a low-cost provider and then check that model’s own batch or cached-input terms if your workload can use them. Discounts and eligibility vary by model and provider. TokenScale compares standard on-demand observations, marks carried values, and uses exact model-specific Batch prices where public terms are established; it never applies one universal discount. Open-weight models on inference hosts are often among the cheapest.

Is TokenScale free?

Yes. TokenScale is free, with no sign-up and no account. It runs entirely in your browser. Nothing you type is sent anywhere or stored.

How often is the pricing updated?

Every night. TokenScale checks provider pricing sources and records what the check could reproduce. A successful observation is verified; otherwise the last known value is carried forward and marked for that date.

4.1 ★★★★☆ AI Critics

Explore ▾

Words

——

Tap to enter a word count ↑ or drag the slider at the top

⚑ report a pricing error

tap to edit

words

≈0 tokens

$0.00

↓ input

$0.00

↑ output

$0.00

= total

Energy lens on — set a word count above

—

of 100W light

Query

—

vs one kettle boil

Kettle Index

—

@ 1M users/day

Carbon

—

gCO₂e per query

Estimated from token count, split ~80% input / 20% output — output decode uses ~3–5× more energy per token than input prefill (Luccioni et al. 2023). Output-heavy tasks (code generation, long-form writing) may use up to 2× more than shown.
Small ~0.0001 / 0.0004 · Mid ~0.0003 / 0.0012 · Flagship ~0.0008 / 0.0035 Wh per 1K in / out tokens.
MoE models (DeepSeek, some open-hosted) adjusted to ~30% of dense equivalent (DeepSeek-V3 report 2024). Range is an illustrative band, not a confidence interval — reflects hardware generation, PUE 1.1–1.6, and batch size (IEA 2024). Carbon uses provider-specific grid mix (IEA 2023): DeepSeek ~580, US providers ~380, global fallback 436 gCO2/kWh. The idea started watching a kettle boil — one litre takes ~0.1 kWh, enough for 30–300 queries like this.

type words · use the slider above · paste text below

Model input Type a number or paste text to count tokens & cost

⚠ Modeled input plus output exceeds this model's context limit. Reduce the request or chunk it.

Cache pricing, model by model

Reuse a prompt or long context? This map shows which exact model routes publish Cache terms.

Select any mark

Model-specific Cache terms will appear here.

Cache pricing coverage for all 66 exact provider and model routes
Provider	① S	② M	③ F

you're on

—

total cost

—

tap a column header to switch tiers & see your cost change

① SMALL

Small · fastest · cheapest

↓

② MID

Mid · balanced · default

↓

③ FLAGSHIP

Flagship · powerful

↓

① ② ③ three tiers · tap a column to switch

—

↓ input

—

↑ output

—

= total

—

↓ input

—

↑ output

—

= total

—

↓ input

—

↑ output

—

= total

—

One model, every tier — this provider has a single model, so all three pricing tiers are identical.

💡

Switch mode to see your savings

Batch and Cache Hit are calculated only for exact models and workloads with current verified official rules.

5 OctNo price moves overnight · 16 of 22 providers verified, the rest carried forwardHistory →

TokenScale

Market

22 providers. Ranked by rate.

— lowest rate

—× most vs least

3 providers compared

Provider pages are authoritative for current published prices. TokenScale checks them nightly and marks any carried observation. Small · Mid · Flagship = each provider’s small, mid & flagship model; the actual model name is always shown.

Tier Provider
Provider · Model	Rate /1M ↑	Cost · vs you

Standard rates · switch mode below

Prices apply current pricing mode · your provider highlighted · drag slider to see costs

Tier
Provider · Model	Rate /1M	Cost · vs you

Standard rates · switch mode below

Fixed A→Z order · rows stay in place as you switch tiers

📖 The Novel Index · checked every night

What it costs to make an AI write a whole novel

The lowest price we find across all 22 providers to generate every one of the 95,356 words in The Hobbit, re-priced each night from each provider's own published rates.

A plain yardstick for how far the price of AI writing has fallen.

Today's lowest

$0.0051

See how we get this →

Tier

Compare provider prices over time

Mid tier · input price per 1M tokens

Providers 4 of 22 selected tap a provider to compare

step lines · list price per 1M tokens · hover a line to spotlight it · click it for one provider · recorded nightly

The Manager's Notepad

October 2026 · Notable price changes

October opens with two provider-directed model handoffs. Z.ai's official migration guide now directs GLM-5.2 users to generally available GLM-5.3, so TokenScale moved the same Zhipu Flagship slot prospectively at the unchanged $1.40/$4.40 standard rate. On October 1, DeepInfra's first-party API explicitly redirected retired Llama 4 Maverick to generally available Gemma 4 31B Turbo. Earlier GLM-5.2 and Maverick history remains intact, and both transitions are labelled as model swaps rather than repricings. Every move, dated →

Oct 5 Both official-source desks Both desks freshly covered all 22 provider routes, 66 Standard tiers, all 66 exact Batch records and all 66 exact Cache records. The independent close recorded 54 verified and twelve grey exact-model holds. At the 11:10 watchdog, two fresh reads of Fireworks' exact DeepSeek V4 Pro (0813) page reproduced the tracked $1.32 input and $3.96 output rate, recovering that unchanged point to verified. The final record verified 55 of 66 standard tiers and carried eleven; Batch and Cache activation rules are unchanged, and no tracked model or list price changed 55 verified · 11 carried provenance recovery

Oct 4 xAI lifecycle + Perplexity source reversal xAI's October 2 release notes say Grok Voice Transcribe 1.0 reached end of life and all calls now redirect to 2.0 at the same price. At the 9:10 watchdog, two fresh Perplexity pricing reads no longer published the three legacy Sonar rate rows seen at 8 PM, while the migration pages still map Sonar and Sonar Pro to Agent API fast and Sonar Reasoning Pro to low. No tracked text model, Standard rate, Batch rule or Cache rule changed. The opening record remains 54 verified and twelve carried 54 verified · 12 carried lifecycle + source reversal

Oct 2 Z.ai guarded successor Z.ai's current migration guide explicitly directs GLM-5.2 users to GLM-5.3. The same Zhipu Flagship slot moved prospectively at the unchanged $1.40 input and $4.40 output per million tokens. Batch remains unestablished. The official table publishes $0.26 cached input and limited-time free cache storage, but incomplete TTL, minimum, eligibility and composition conditions keep public Cache calculation unestablished. The opening desk verified 56 of 66 standard tiers and carried ten exact-model holds grey 56 verified · 10 carried guarded successor

Oct 1 DeepInfra successor and AWS watchdog recovery Both desks freshly covered all 22 routes. The 11:10 watchdog restored AWS Claude Haiku 4.5 at $1/$5 and Llama 4 Maverick at $0.24/$0.97 to verified after fresh rendered Ohio panels reproduced both exact rows; Claude Sonnet 5 remains absent and carried grey. The final record verifies 56 of 66 standard tiers. DeepInfra's exact retired Maverick record still names google/gemma-4-31B-it-turbo as its replacement, and the successor remains $0.09 input and $0.34 output per million tokens. Batch remains 21 supported, five unsupported and 40 unestablished. Cache remains six supported, three explicitly unavailable, 52 unestablished and five blocked 56 verified · 10 carried guarded successor

Data through 5 October 2026 · verification detail loading · Bilton Projects

September 2026 · Notable price changes

On 29 September, OpenRouter's live DeepSeek V3 0324 route changed from DeepInfra at $0.24/$0.90 to SiliconFlow at $0.25/$1, while OpenAI launched GPT-6.1 Sol at $2/$10 without naming it as GPT-6 Sol's successor. TokenScale recorded the same-model route repricing and held the lineup choice for Will. On 25 September, at Will's direction, TokenScale moved three tracked slots to this week's cheaper frontier launches: OpenAI Small to GPT-6 Luna ($0.10/$0.50), OpenAI Flagship to GPT-6 Sol ($2/$10) and Anthropic Flagship to Claude Opus 5.5 ($4/$20). Each earlier model keeps its own dated history. September opens with frontier launches and two guarded successors. Anthropic explicitly named Claude Fable 5.1 as the generally available successor to Claude Fable 5, so TokenScale moved the Pro slot prospectively on 3 September at the same $10/$50 standard rate. On 10 September, DeepSeek's exact Lite alias began automatically routing legacy V4 Flash names to generally available V4.1 Flash, so TokenScale moved that slot prospectively to the new $0.30/$1.20 peak rate and exact $0.15/$0.60 off-peak schedule. On 11 September, DeepSeek explicitly withdrew its planned 14 September V4 Pro redirect and said API service will continue with billing unchanged. The 27 September watchdog recovered the missing opening desk and verified 60 of 66 standard tiers, with six exact-model holds visibly carried. DeepInfra scheduled tracked Llama 4 Maverick to redirect on 1 October. Every move, dated →

Sep 30 Opening desk and October 1 redirect watch The opening desk freshly covered all 22 routes and verified 56 of 66 standard tiers. AWS Haiku 4.5 and Llama 4 Maverick returned to verified after current Ohio rows reproduced, while Sonnet 5 remained absent. The ten exact-model carries stay visible. Batch remains 21 supported, five unsupported and 40 unestablished; Cache remains six supported, three explicitly unavailable, 52 unestablished and five blocked. DeepInfra still serves tracked Llama 4 Maverick at $0.20/$0.80 before its scheduled October 1 automatic redirect, so no successor was applied early 56 verified · 10 carried opening pass

Sep 29 OpenRouter repricing and GPT-6.1 Sol launch The opening desk covered all 22 routes and verified 54 of 66 standard tiers. OpenRouter's exact endpoint API now exposes tracked DeepSeek V3 0324 through SiliconFlow at $0.25/$1, replacing the former cheapest DeepInfra route at $0.24/$0.90; the same-model change is recorded prospectively. OpenAI launched generally available GPT-6.1 Sol at $2/$10 with exact Batch and Cache terms, but published no successor mapping from tracked GPT-6 Sol, so the lineup remains unchanged pending Will 54 verified · 12 carried route repriced

Sep 28 Anthropic successor and Perplexity migration hold Both official-source desks covered all 22 routes, and the independent close verified 56 of 66 standard tiers. Anthropic's first-party migration guide directs Claude Sonnet 5 users to generally available Claude Sonnet 5.5 at unchanged $2/$10 Standard, $1/$5 Batch and exact model-specific Cache terms, so TokenScale applied the same-provider Mid-tier guarded successor prospectively and preserved every earlier Sonnet 5 point. Perplexity's closing reads reproduced all three rates, but its migration changes service scope and collapses distinct rows into Agent API routing modes, so all three tiers remain visibly carried pending Will's lineup decision 56 verified · 10 carried guarded successor

Sep 27 Watchdog recovery and lifecycle watch The 9:10 watchdog found the 8 PM receipt missing and ran the full 22-route recovery. Standard remains 60 verified and six carried, Batch remains 21 supported, five unsupported and 40 unestablished, and Cache remains six supported, three explicitly unavailable, 52 unestablished and five blocked. Together's exact model pages now reproduce its two held rates, but its distinct current serverless catalog still omits both models, so the conflict remains grey. DeepInfra scheduled tracked Llama 4 Maverick to redirect to Gemma 4 31B on 1 October, while OpenRouter says tracked DeepSeek V3 0324 goes away on 29 September. No future lineup change was applied early 60 verified · 6 carried recovery recorded

Sep 25 Three tracked model changes At Will's direction, OpenAI Small moved from GPT-5.6 Luna to GPT-6 Luna at $0.10/$0.50, OpenAI Flagship moved from GPT-5.6 Sol to GPT-6 Sol at $2/$10, and Anthropic Flagship moved from Claude Fable 5.1 to Claude Opus 5.5 at $4/$20. The earlier models keep their own dated history. The closing desk covered all 22 official routes. At 11:10, two fresh cache-busted AWS Ohio renders reproduced Claude Haiku 4.5 at $1/$5 with its Batch and Cache columns, and the distinct current AWS Batch model list named Haiku 4.5. Claude Sonnet 5 remains grey; Perplexity's September 27 Sonar support sentence remains withdrawn 60 verified · 6 carried model swaps

Sep 24 Closing source reversals Both official-source desks covered all 22 routes. The independent close again verified 60 of 66 Standard tiers and carried six exact-model holds. Perplexity's current migration page restored the statement that Sonar will be supported until September 27, 2026. DeepInfra's main table and OpenRouter's model list omitted one tracked row, but fresh exact-model pages reproduced DeepInfra Llama 4 Maverick at $0.20/$0.80 and OpenRouter DeepSeek V3 0324 at $0.24/$0.90 with $0.135 cached input. No tracked rate, lineup, promotion, Batch rule or Cache rule changed 60 verified · 6 carried closing recorded

Sep 23 Launch and closing source record Anthropic launched generally available Claude Opus 5.5 at $4 input and $20 output per million tokens with a 1M-token context window. It does not explicitly replace tracked Claude Fable 5.1, so no lineup changed. Both official-source desks covered all 22 routes. The independent close reproduced Claude Haiku 4.5 at $1/$5 and Llama 4 Maverick at $0.24/$0.97 in Ohio; Sonnet 5 remained absent and grey with the five standing exact-model holds. No tracked rate, Batch rule or Cache rule changed 60 verified · 6 carried watchdog recovery

Sep 22 Closing source record Both official-source desks covered all 22 routes. AWS's closing rendered Anthropic table again reproduced Claude Haiku 4.5 at $1/$5 but omitted tracked Claude Sonnet 5; its Meta table reproduced Llama 4 Maverick at $0.24/$0.97. TokenScale therefore carries Sonnet 5's prior $2/$10 rate grey. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5 remain the other exact-route holds. No tracked rate, lineup, Batch rule or Cache rule changed 60 verified · 6 carried close recorded

Sep 21 Launch and source hold Both official-source desks covered all 22 routes. The independent close reproduced AWS Claude Haiku 4.5 at $1/$5, Claude Sonnet 5 at $2/$10 and Llama 4 Maverick at $0.24/$0.97 in US East (Ohio), resolving the opening source gap without changing a rate. xAI launched generally available Grok 4.7 with a 500K context window at $2 input, $0.50 cached input and $6 output per million tokens below 200K context, with $4/$1/$12 at or above 200K. Tracked Grok 4.5 remains separately listed, so no lineup change was made. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5 remain the five exact-route holds 61 verified · 5 carried closing recorded

Sep 20 Independent close Both official-source desks reopened all 22 routes. OpenRouter DeepSeek V3 0324 again reproduced on one DeepInfra route at $0.24/$0.90 with $0.135 cached input. At the 11:10 PM watchdog, two fresh canonical pricing reads, the current Sonar catalog and the exact Sonar, Sonar Pro and Sonar Reasoning Pro pages all omitted the September 27 support sentence. TokenScale withdrew only that lifecycle claim; the unchanged Sonar $1/$1, Sonar Reasoning Pro $2/$8 and Sonar Pro $3/$15 rates remain verified. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain the exact-route holds 61 verified · 5 carried close recorded

Sep 19 Closing source reversal All 22 official routes were freshly reopened. OpenRouter DeepSeek V3 0324 again reproduced on one DeepInfra route at $0.24/$0.90 with $0.135 cached input. Perplexity still publishes Sonar at $1/$1, Sonar Reasoning Pro at $2/$8 and Sonar Pro at $3/$15, but the closing page no longer publishes the September 27 support sentence seen at 8 PM. The closing read controls, so TokenScale withdraws only that lifecycle claim. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain the exact-route holds 61 verified · 5 carried close

Sep 18 Closing source reversals All 22 official routes were freshly reopened. The 9:10 PM watchdog and independent 10 PM close reproduced OpenRouter DeepSeek V3 0324 on one DeepInfra route at the unchanged $0.24/$0.90 pair, with $0.135 cached input and a distinct current model-list confirmation. Perplexity's closing page again states that Sonar will be supported until September 27, restoring that lifecycle fact. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain the exact-route holds 61 verified · 5 carried 2 reversals closed

Sep 17 Opening source reversal All 22 official routes were freshly reopened. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain visible source holds. Perplexity's current first-party page still publishes Sonar at $1/$1, Sonar Reasoning Pro at $2/$8 and Sonar Pro at $3/$15, but not the earlier September 27 support sentence. TokenScale has withdrawn only that lifecycle claim pending the independent 10 PM close 61 verified · 5 carried claim withdrawn

Sep 16 First source pass Both official-source desks freshly reopened all 22 routes. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain visible source holds. Perplexity's independent close again states that Sonar support continues through September 27, confirming the restored lifecycle fact; all three token rates remain unchanged 61 verified · 5 carried source reversal

Sep 15 First-desk watchdog All 22 official routes were freshly reopened. The watchdog then reproduced OpenRouter DeepSeek V3 0324 twice on one DeepInfra route at the unchanged $0.24/$0.90 pair, upgrading that same-day point. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain visible source holds. Google also launched two generally available Gemini 3.8 Live audio-to-audio models; they do not replace the tracked text lineup 61 verified · 5 carried source reversal

Sep 14 First source pass All 22 official routes were freshly reopened. Sixty exact Standard rows reproduced unchanged. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Moonshot Kimi K2.5 and OpenRouter's split DeepSeek V3 route remain visible source holds. Perplexity's three rates remain unchanged and its withdrawn September 27 lifecycle sentence remains absent 60 verified · 6 carried source pass

Sep 13 Perplexity closing read The current first-party page still publishes Sonar at $1/$1, Sonar Reasoning Pro at $2/$8 and Sonar Pro at $3/$15, but no longer publishes the September 27 support sentence seen earlier the same day. The closing read controls, so TokenScale withdraws only that lifecycle claim rates unchanged · claim withdrawn source reversal

Sep 13 First-desk watchdog All 22 official routes were freshly reopened. Perplexity again states that Sonar will be supported until September 27, restoring the mutable lifecycle claim subject to the 10 PM close. The watchdog then reproduced AWS Haiku 4.5 and Sonnet 5 plus Meta Muse Spark 1.2 twice on rendered official pricing pages, with distinct current model pages confirming each tracked model. Six exact-route holds remain 60 verified · 6 carried source pass

Sep 11 DeepSeek DeepSeek's current pricing page now explicitly says V4 Pro API service will continue after September 14 with billing unchanged, withdrawing the earlier planned redirect. Both tracked V4 Pro slots and prices remain unchanged planned redirect withdrawn source reversal

Sep 11 First source pass All 22 official routes were freshly reopened. OpenRouter's former single DeepSeek V3 route disappeared; StreamLake owns the lower input while DeepInfra owns the lower output, so TokenScale carries the prior pair instead of mixing routes. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Moonshot Kimi K2.5 remain exact-model holds 60 verified · 6 carried source pass

Sep 10 DeepSeek DeepSeek's exact deepseek-flash route began serving V4.1 Flash and automatically redirecting legacy V4 Flash names. The generally available same-route successor moved TokenScale's Lite slot prospectively to $0.30 input and $1.20 output at peak, with exact $0.15/$0.60 off-peak rates. DeepSeek withdrew its separately planned V4 Pro redirect on September 11 $0.44/$1.32 → $0.30/$1.20 guarded successor

Sep 10 Closing watchdog MiniMax's current API overview explicitly lists MiniMax M2 as a supported Large Language Model API model. Two fresh rendered pricing reads reproduced its unchanged $0.30 input and $1.20 output rate plus exact $0.03 cache-read, $0.375 cache-write and five-minute explicit-cache terms 61 verified · 5 carried source recovery

Sep 10 First source pass All 22 official routes were freshly reopened. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Moonshot Kimi K2.5 and legacy MiniMax M2 remain exact-model source holds. DeepSeek V4.1 Flash is the only tracked Standard change 60 verified · 6 carried source pass

Sep 9 First source pass All 22 official routes were freshly reopened. AWS Bedrock's three tracked Standard rows reproduced in the rendered regional tables, while Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Moonshot Kimi K2.5 and legacy MiniMax M2 remain exact-model source holds. Perplexity's three token-rate rows verified, but its current page no longer reproduced the earlier September 27 Sonar-support sentence, so that mutable lifecycle claim is withdrawn pending the 10 PM close 60 verified · 6 carried source pass

Sep 6 First source pass All 22 official routes were freshly reopened. Perplexity again published the September 27 Sonar-support sentence alongside unchanged Sonar $1/$1, Sonar Reasoning Pro $2/$8 and Sonar Pro $3/$15 rates. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Moonshot Kimi K2.5 and legacy MiniMax M2 remain exact-model source holds 60 verified · 6 carried source pass

Sep 5 OpenAI GPT-6 Astra's current model page now lists Tier 1 limits and the current catalog recommends it as OpenAI's flagship for general use. TokenScale has not changed its tracked lineup; Will owns that model-choice decision $10/$50 · lineup held launch update

Sep 5 First source pass All 22 official routes were freshly reopened. The 8 PM view omitted Cohere Command R+, but two fresh 9:10 watchdog reads reproduced its exact $2.50/$10 row and the current model table still marked it Live. Together GPT-OSS 20B and Kimi K2.6, Cerebras ZAI-GLM-4.7, Fireworks GPT-OSS 20B, retired Moonshot Kimi K2.5 and legacy MiniMax M2 remain exact-model source holds. The missing 4 September shipping points were restored explicitly as carried history 60 verified · 6 carried source pass

Sep 5 Closing source reversal Perplexity's 10 PM first-party pricing page reproduced Sonar at $1/$1, Sonar Reasoning Pro at $2/$8 and Sonar Pro at $3/$15, but no longer reproduced the previously visible September 27 support sentence. That lifecycle claim has been withdrawn from current copy; no tracked model or price changed prices unchanged · claim withdrawn close

Sep 4 Closing source reversal Perplexity's first-party pricing page again stated that Sonar support continues through September 27, 2026, superseding the 8 PM omission. MiniMax M2 also remains grey: its exact Standard and Cache rows remained present in static first-party pricing HTML, but the rendered current catalog kept M2 under Legacy Models and did not establish current callable eligibility 60 verified · 6 carried close

Sep 4 Watchdog source reversal MiniMax M2's exact $0.30 input, $1.20 output, $0.03 cache-read and $0.375 cache-write rows returned on two fresh first-party pricing reads. A distinct current catalog listed M2 only under Legacy Models, so its Standard point stayed grey and Cache stayed blocked pending the independent 10 PM close source restored · still grey source reversal

Sep 4 First source pass All 22 official routes were freshly reopened, with 60 of 66 standard tiers verified. Six exact-model rows remained held: Together GPT-OSS 20B and Kimi K2.6, Cerebras ZAI-GLM-4.7, Fireworks GPT-OSS 20B, retired Moonshot Kimi K2.5 and MiniMax M2 60 verified · 6 carried source pass

Sep 3 Anthropic Claude Fable 5.1 replaced Claude Fable 5 in the Pro slot after Anthropic explicitly named it the generally available successor. Standard and Batch rates remain $10/$50 and $5/$25; cache read changes from $1 to $0.25 per million tokens successor · $10/$50 model swap

Sep 3 OpenAI GPT-6 Astra launched at $10 input and $50 output per million tokens with $1 cached input. Its launch-day page limited access to Trusted Access; a 5 September follow-up above records the current broader catalog wording launch · lineup held launch

Sep 3 First source pass All 22 official routes were freshly reopened. Together GPT-OSS 20B and Kimi K2.6, Cerebras ZAI-GLM-4.7, Fireworks GPT-OSS 20B and retired Moonshot Kimi K2.5 remain exact-model source holds; no temporary promotion entered canonical history 61 verified · 5 carried source pass

Sep 2 Google Gemini 3.8 Flash is now available with a temporary $0.75/$3.75 rate through December 31 and a $1.50/$7.50 list rate from January 1, matching the still-listed tracked Gemini 3.6 Flash rates launch · lineup held launch

Sep 2 Anthropic Claude Fable 5.1 now appears in first-party pricing at $10 input and $50 output per million tokens. Tracked Claude Fable 5 remains listed at the same standard rate $10/$50 · lineup held launch

Sep 2 First source pass All 22 official routes were reopened twice. At close, Together GPT-OSS 20B joined Together Kimi K2.6, Cerebras ZAI-GLM-4.7, Fireworks GPT-OSS 20B and Moonshot Kimi K2.5 as exact-model source holds; no replacement or temporary promotion entered canonical history 61 verified · 5 carried source pass

Data through 30 September 2026 · 56 of 66 tiers verified at month close; 10 carried · Bilton Projects

August 2026 · Notable price changes

Aug 31Moonshot KimiMoonshot's current Model List says tracked Kimi K2.5 was retired today and calls now return HTTP 404. The prior $0.60/$3 point remains grey; K2.6, K2.7 Code and K3 are not automatic substitutes62 verified · 4 carriedretired · held

August is mostly an audit month. On 18 August, TokenScale adopted a conservative rule for time-window pricing: the comparison uses the highest standard on-demand rate and labels it as peak, while publishing the exact lower schedule. DeepSeek now verifies at V4 Flash peak $0.44/$1.32 and V4 Pro peak $1.32/$3.96; both are 50% lower off-peak outside 01:00–04:00 and 06:00–10:00 UTC. Groq's retired developer-plan Llama rows were replaced with its current production GPT-OSS 20B and 120B models. On 21 August, OpenAI began showing a clearly labelled $4/$20 Sol promotion while its $5/$30 list series stayed unchanged. Z.ai launched GLM-5.3-Flash on 26 August with a temporary 50% rate reduction through September 9; TokenScale has not changed its tracked GLM lineup. At the 30 August first desk, 63 of 66 tiers verified. Together Kimi K2.6 and Cerebras ZAI-GLM-4.7 remain held grey, and Fireworks GPT-OSS 20B joined them after its official rendered page said serverless is not supported and the static serverless table exposed no current rate. OpenRouter's exact DeepSeek V3 route reproduced $0.24/$0.90 with $0.135 cached input. Perplexity's three rates still verify, and its current page states that Sonar support continues through September 27. Every move, dated →

Aug 30Fireworks AIThe tracked GPT-OSS 20B rendered model page now says Serverless: Not supported, while the serverless pricing table exposes no current row. The prior $0.07/$0.30 remains grey pending retry or Will's replacement decision63 verified · 3 carriedsource hold

Aug 26Z.aiGLM-5.3-Flash launched with a limited-time 50% reduction through September 9; the tracked GLM lineup remains unchanged pending Will's decisionpromo $0.075/$0.25 · list $0.15/$0.50launch

Aug 29 Fireworks AI The exact DeepSeek V4 Pro (0813) standard row now publishes $1.32 input, $0.044 cached input and $3.96 output per million tokens $1.74/$3.48 → $1.32/$3.96 repricing

Aug 29 First source pass OpenRouter DeepSeek V3 returned to one reproducible DeepInfra route at $0.24/$0.90. Together Kimi K2.6 and Cerebras ZAI-GLM-4.7 remain source holds pending the independent closing pass or Will's replacement decision 64 verified · 2 carried source pass

Aug 28 Closing source pass Together Kimi K2.6, Cerebras ZAI-GLM-4.7 and OpenRouter DeepSeek V3 remain source holds. Perplexity's closing read again states that Sonar support continues through September 27, superseding the earlier same-evening omission 63 verified · 3 carried carry

Aug 25 Full source pass Together Kimi K2.6 and Cerebras ZAI-GLM-4.7 remain source holds. Perplexity's closing read again states that Sonar support continues through September 27, superseding the earlier same-day omission 64 verified · 2 carried carry

Aug 23 Recovery source holds A recovered full pass found Together still omitting tracked Kimi K2.6 and Cerebras still omitting ZAI-GLM-4.7. Both prior rates remain visibly grey pending a stable current row or Will's replacement decision 64 verified · 2 carried carry

Aug 22 Closing source holds Together's tracked Kimi K2.6 row appeared at 8 PM but was absent at the independent 10 PM close. Cerebras still omits ZAI-GLM-4.7. Both prior rates remain visibly grey pending a stable current row or Will's replacement decision 64 verified · 2 carried carry

Aug 21 OpenAI GPT-5.6 Sol now shows a temporary promotional rate through at least 21 November; the permanent list-price history remains unchanged promo $4/$20 · list $5/$30 promo

Aug 21 Source holds Together's current inference table omits tracked Kimi K2.6, while Cerebras still omits ZAI-GLM-4.7. Both prior rates remain visibly grey pending a current row or Will's replacement decision 64 verified · 2 carried carry

Aug 18 Closing source repair MiniMax M2 verified at $0.30/$1.20 from official static HTML; Z.ai's public pay-as-you-go table now lists GLM-5.3 at $1.40/$4.40, with the lineup held for Will 60 verified · 6 carried source fix

Aug 18 DeepSeek The board now uses the maximum standard on-demand rate for time-window schedules, labels it peak and publishes the lower window without averaging Flash peak $0.44/$1.32 · Pro peak $1.32/$3.96 · off-peak 50% lower repricing

Aug 18 Groq Retired Llama developer rows replaced with the two models in Groq's current public Production Models table; 120B fills both larger tiers rather than using a preview model Small GPT-OSS 20B $0.075/$0.30 · Mid + Flagship GPT-OSS 120B $0.15/$0.60 model swap

Aug 17 Closing correction DeepSeek's article returned to peak and off-peak V4 rates between desks; Groq and Cerebras rows were absent, and MiniMax's rendered page hid M2. The later static-source recovery does not recolor 17 August 59 verified · 7 carried correction

Aug 16 Alibaba Qwen QwenCloud's public model page now publishes exact qwen3.8-max pricing; Qwen3.7-Max remains tracked pending Will's lineup decision $2 in · $6 out watch

Aug 16 DeepSeek Peak and off-peak V4 pricing is now live; three prior flat-rate slots are carried grey until the daily board has a time-window rule Flash $0.22/$0.66 to $0.44/$1.32 · Pro $0.66/$1.98 to $1.32/$3.96 carry

Aug 16 Groq Llama 3.1 8B and Llama 3.3 70B reached their developer-plan shutdown date, but Groq still lists both as Production Models and exempts committed-spend enterprise users replacement lineup held for Will watch

Aug 16 TokenScale Correction: Sonnet 5's $2/$10 introductory price became the standard price, but the amount customers paid did not fall on 12 Aug $2/$10 paid before and after correction

Aug 16 TokenScale Correction to the event date: OpenAI's dated announcement makes the Luna and Terra cuts effective 30 Jul, verified by TokenScale on 12 Aug moved to July record correction

Aug 14 DeepSeek Peak and off-peak V4 rates take effect 16 August at 16:00 UTC; today's canonical rates remain unchanged Flash $0.22/$0.66 to $0.44/$1.32 · Pro $0.66/$1.98 to $1.32/$3.96 watch

Aug 14 OpenRouter The previous DeepSeek V3 route disappeared; current routes split the cheapest input and output, so the mid tier is held grey pending a route rule no fabricated price pair carry

Aug 13 Google Official Gemini API pricing now lists Gemini 3.7 Flash; Gemini 3.6 Flash remains tracked pending Will's lineup decision temporary $0.75/$3.75 through 31 Dec watch

Aug 12 OpenRouter DeepSeek V3: cheapest live route fell again on the fresh evening API read $0.2574/$1.0287 → $0.24/$0.90 cut

Aug 12 Launch watch Official pages show Grok 4.6 and Kimi K2.7 Code. Alibaba refers to a Qwen3.8 family, but the exact reported Max Preview row and public rate are not reproducible; all tracked lineups remain held pending Will no automatic model swap watch

Aug 11 OpenAI Superseded correction: a model-page conflict led us to withdraw the Luna/Terra claim. OpenAI's dated 30 Jul announcement later resolved the event date and rates retained as audit trail superseded

Aug 11 AWS Bedrock Pro slot: EOL Llama 3.1 405B replaced by active Llama 4 Maverick 17B Instruct; Haiku 4.5 and Sonnet 5 are now pinned to exact SKUs $2.40/$2.40 → $0.24/$0.97 model swap

Aug 11 Cohere Label correction: Command R7B 12-2024, Command R 08-2024 and Command R+ 08-2024 are live; the deprecated undated aliases are gone rates unchanged pinned

Aug 10 xAI Flagship slot: Grok 4.20 → Grok 4.5 (xAI's featured model since 9 Jul). 4.20 still sold at $1.25/$2.50 $1.25/$2.50 → $2/$6 model swap

Aug 10 Meta Provider now tracks Muse Spark 1.2, Meta's paid API. Llama rows retired to the graveyard; history before 10 Aug is Llama pricing $0.15/$0.60 → $1.25/$4.25 model swap

Aug 10 OpenRouter Mid (DeepSeek V3): cheapest live route repriced. Route basis: cheapest live deepseek-chat route, read from OpenRouter's API $0.2002/$0.8001 → $0.2574/$1.0287 +29%

Aug 5 Verify ×22 Late pass: 13 read clean at source, 9 held as carried. The later Luna/Terra rate claim was withdrawn on 11 Aug 13 stamped · 9 carried verify

Data through 31 August 2026 · 62 of 66 tiers verified at month close; 4 carried · Bilton Projects

July 2026 · Notable price changes

The busiest month this board has tracked. It opened with a reckoning, Claude Fable 5 returned and re-audited all 66 tiers in one sweep (40 exact, 26 fixed), four labs shipped inside three weeks, and OpenAI cut Luna 80% and Terra 20% on 30 July. Every move, dated →

Jul 30 OpenAI GPT-5.6 Luna, effective 30 Jul and verified 12 Aug $1/$6 → $0.20/$1.20 −80%

Jul 30 OpenAI GPT-5.6 Terra, effective 30 Jul and verified 12 Aug $2.50/$15 → $2/$12 −20%

Jul 28 Corrections ×2 Groq flagship tracked a model dead since 15 Apr · Qwen read the wrong region Groq $1.00→$0.15 in · Qwen $0.165→$0.25 in corrections

Jul 24 Anthropic Claude Opus 5 launch (all-models; Fable 5 keeps the flagship slot) $5/$25 list · 1M context new

Jul 21 Google Gemini Mid (Flash 3.6 · a new generation that costs less) $9.00→$7.50 out · input flat at $1.50 −17%

Jul 16 Moonshot Kimi Flagship (Kimi K3 · 1M context, priciest Chinese-lab launch yet) $1.90→$3.00 in · $8.00→$15.00 out model swap

Jul 9 OpenAI GPT-5.6 GA (Sol · Terra · Luna take all three tiers) lite $0.75→$1.00 in · $4.50→$6.00 out model swap

Jul 1 Anthropic Pro (Fable 5 returns, retakes the slot; Opus 4.8 stays listed) $5→$10 in · $25→$50 out it's back

Jul 1 xAI Grok Lite (4.1 Fast retired, old slugs bill 6×) $0.20→$1.00 in · $0.50→$2.00 out model swap

Jul 1 Mistral AI Flagship (Large 3, now cheaper than Medium) $2.00→$0.50 in · $6.00→$1.50 out model swap

Jul 1 Audit ×22 Phantoms out: K2.6 Turbo · M2.7 Pro · Behemoth 40/66 exact · 26 tiers fixed audit

July 2026 · verified at month close · Bilton Projects

How verification works Official sources · checked nightly Open +Close −

Grey is not hidden uncertainty. It marks the exact tier and date where the last known price was carried forward because the official rate could not be reproduced that day.

Black: we read the provider's official pricing page on that date.

Grey: the last known tier price was carried forward. It was not re-checked that day.

Check a provider below ↓ See every missed check → Read the audit journal →

The market, measured: every list-price move since May 2026, counted and classified, medians by tier and the full spread. The State of AI Pricing read the report →

📈 Tap a sparkline · use ← → to browse all 22 charts →

👆 Tap any row to inspect its source, checked date and grey days

✓ Verified monthly by Bilton Projects · October 2026

⚡ Compare providers: 💻 Shift+click rows · 📱 hold ½ sec — build a list, export CSV

your task

One growing conversation Each reply re-reads the whole history ▾

Unlike daily volume above, this follows one conversation. Each reply re-reads its growing history.

one conversation · each reply re-reads the growing history

THE COMPOUNDING EFFECT

Every reply re-reads
the entire conversation

Each time the AI responds, it processes all previous messages from scratch — not just your question. By reply 5, you're paying for 15 context re-reads. By reply 10, context dominates your bill.

↓ drag the reply slider to watch costs stack in real time

context re-read (grows) output written (fixed)

— in : out ratio —

per-call cost × volume = real budget · see the scale

PRODUCTION REALITY CHECK

Fractions of a cent become
real budget at scale

A $0.0001 per-call cost looks free. At 1,000 calls/day it's $3/month — manageable. At 100K calls/day it's $300/month. Set your actual volume and discover where AI fits in your budget.

↓ drag the sliders to reveal your production number

API calls per day production volume

1,000 calls +10K +100K

Output words per call avg. reply length

300 words +250 +1000

Per call

—

enter a word count

Per day

—

Per month

—

30 days

Per year

—

365 days

Common questions ∨

What does an AI API call actually cost?

Why does output price matter more than input?

How much do prices differ between providers?

What is the cheapest way to run a large workload?

Provider name

provider type

Everything you need to understand AI model pricing, providers and how to choose.

⚠ Report a pricing error

Help keep TokenScale accurate

What's wrong?

Source URL (optional but helpful)

Your email (optional — only if you want a reply)

Help me choose

Question 1 of 5

What do you mainly want to use AI for?

Select all that apply

Our recommendation

—

Also worth considering

Learn more

Tap to close

AI pricing, measured
for your workload.

The whole idea, in plain English

Pick something you'd write

Now pick an AI

Your price

Full calculator

The whole idea, in plain English

Welcome to the calculator

AI pricing, measuredfor your workload.

The whole idea, in plain English

1Pick something you'd write

2Now pick an AI

3Your price

AI pricing, measured
for your workload.

Pick something you'd write

Now pick an AI

Your price