Every product update and price movement — explained in words, not API docs. Because a 400% price increase means more when you see it as a novel going from $0.35 to $1.33.
This page is a dated record. Prices quoted inside an entry are the prices as recorded on that entry’s date, not necessarily today’s rates. Current prices live on the comparison table, checked nightly with carried observations marked.
Google released generally available Gemini 4 Neon at $3 input and $15 output per million tokens, with $0.15 cached input, implicit-only caching, a 1M-token context window and four thinking levels.
The current first-party pages do not uniquely map Neon to tracked Flash-Lite 3.5, Flash 3.6 or Pro 3.1. TokenScale records the launch and leaves the lineup decision with Will.
OpenAI released GPT-6.2 Astra at $8 input and $40 output per million tokens and GPT-6.2 Sol at $2 input and $10 output, with separate cached-input, long-context and priority-processing rules.
The current first-party pages do not uniquely map either launch to one tracked TokenScale slot. TokenScale records the launch and leaves the lineup decision with Will.
OpenAI added two-hour and eight-hour background completion windows for selected models. Anthropic opened linked billing in beta. Google added voice editing to Gemini 3.5 Transcribe. Together listed untracked Elorian Glimpse 1.0 at $0.60 input, $0.06 cached input and $3 output per million tokens.
The four facts crossed the mechanical source-watch threshold. They do not change a tracked TokenScale Standard price, exact Batch rule, Cache field, model slot or Journal premise.
Alibaba launched generally available Qwen4-Max at $1.50 input and $4.50 output per million tokens on its international Model Studio service, with explicit cache-write and cache-read terms.
Alibaba calls it a next-generation flagship but does not uniquely map it to tracked Qwen3.7-Max. TokenScale records the launch without inferring a lineup change.
OpenAI added the free, approval-gated omni-moderation-csam-latest specialist classifier. Anthropic extended text watermarking to current Opus 4.6, Opus 4.5, Sonnet 4.5 and tracked Haiku 4.5. Mistral documented low, medium and high reasoning-effort modes for Large 4, which the closing desk confirmed is now generally available.
The three facts crossed the mechanical source-watch threshold. They do not change a tracked TokenScale Standard price, exact Batch rule, Cache field, model slot or Journal premise.
Google's current first-party release notes list generally available Gemini 4 Hydrogen at $1.50 input and $7.50 output per million tokens. The release points migration from Gemini 3.8 Flash, not TokenScale's tracked Gemini 3.6 Flash or Gemini 3.1 Pro rows, so TokenScale records the launch without inferring a successor.
OpenAI's GPT Image 2 lifecycle notice points users to GPT Image 2.6 variants. Alibaba's international pricing page shows an untracked Qwen4-Max-Preview Token Plan banner. Google released the specialist Gemini Decision 1 and Decisions API. Anthropic launched self-hosted Claude Managed Agents for organizations running the full service on their own infrastructure.
The oldest item completed its four-night watch and the cluster exceeded the three-item threshold. None changes a tracked TokenScale Standard price, exact Batch rule, Cache field, model slot or Journal premise.
OpenAI released the untracked GPT-Image-2.6 Sunburst and Flare image models. Google added voice remixing for Gemini 3.8 TTS. Kimi K3.1 Pro gained new official hosting listings across Together, Fireworks, DeepInfra and Alibaba Model Studio. Cohere's current model page now maps tracked Command models to OpenRouter. OpenAI also extended the GPT-5.6 Sol promotion wording through 30 November; that model left the tracked slot on 25 September.
The five facts crossed the mechanical source-watch threshold. They do not change a tracked TokenScale Standard price, exact Batch rule, Cache field, model slot or Journal premise.
xAI retired the untracked grok-imagine-image-quality route and redirects it to grok-imagine-image-2.0 with low quality. Anthropic added beta Messages API threads backed by short-lived server-side containers. Anthropic also documented text watermarking for Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6.
The three facts crossed the mechanical source-watch threshold. They do not change a tracked TokenScale Standard price, exact Batch rule, Cache field, model slot or Journal premise.
Groq's current public model pages again reproduce GPT-OSS 20B at $0.075 input and $0.30 output per million tokens and GPT-OSS 120B at $0.15/$0.60. All three tracked Standard rows return to verified. Batch reproduces at half the Standard rates. Prompt caching also reproduces at half input with automatic two-hour volatile storage, but TokenScale keeps Cache calculation unestablished because Groq publishes only a model-dependent 128 to 1,024 token minimum range.
The same pass records OpenAI's GPT-6 Astra legal configuration, its 10 percent FedRAMP uplift starting November 1, and xAI output watermarking for tracked Grok text models starting November 7. These are factual API conditions; none changes a canonical TokenScale Standard rate or model slot.
Kimi K2.6 is again published at $0.95 input, $4 output and $0.16 cached input per million tokens, with exact Batch rates of $0.57/$2.40 and $0.10 cached input. Two fresh pricing reads and the distinct current model catalog reproduced the row, so TokenScale restores both K2.6 roles to verified without changing their values or the lineup. K3's exact Standard, Batch and cache rates also reproduced.
Mistral's current model and pricing surfaces no longer publish exact Medium 3.5 or Large 3 rows and list both deprecated, with Large 4 named as successor. Large 4 remains Public Preview and one candidate maps to two TokenScale roles, so Medium 3.5 and Large 3 stay visibly carried at their prior values pending a lineup decision.
The same pass records OpenAI GPT-6 Rosalind trusted access, Anthropic's Opus 4.8 fast-mode deprecation, and Google Gemini Computer Use general availability as a three-item source-watch flush. None changes a tracked price, exact Batch rule, Cache activation, model slot or Journal premise.
xAI added custom moderation controls. OpenAI added an included API-credit benefit for eligible ChatGPT Pro and Business subscribers. Anthropic launched the Usage and Cost Analytics API in beta while keeping its prior Usage and Cost API and Enterprise reporting endpoints supported as legacy surfaces. Anthropic also raised Messages API and Token Counting API request limits from 32 MB to 128 MB.
The four-item queue crossed the mechanical source-watch threshold. TokenScale records the factual cluster without projecting a Standard price move, exact Batch rule, Cache field, model slot or Journal premise. No credit claim, account action or paid action was taken.
OpenAI launched GPT-6.1 Terra with a 1.05-million-token context window at $1 input and $5 output per million tokens. Its published prompt-cache terms are $0.05 per million cached-input tokens, $1.25 for five-minute cache writes and $2.50 for extended cache writes.
The provider did not direct users from TokenScale's tracked GPT-5.6 Terra to the new model, and the tracked model's exact page still publishes its current $2/$12 Standard row. TokenScale records the Class A launch without rewriting history or making a model-choice decision; Will owns the prospective Mid lineup.
OpenAI's 23 October deprecation notice scheduled shutdowns for legacy untracked API snapshots. Together now applies a 25% premium to geo-specific inference while its tracked public serverless rows remain unchanged. Cohere introduced the Embed 5 family, an embedding-model addition outside TokenScale's tracked Command text lineup.
The two held items reached their four-night limit as the Cohere fact brought the queue to three. TokenScale records the factual cluster without projecting a Standard price move, Batch rule, Cache field or successor into any tracked slot.
Kimi now lists generally available K3.1 Pro at $2 input and $10 output per million tokens and K3.1 Pro Highspeed at $4/$20. The same public platform surface continues to list TokenScale's tracked K3 at its unchanged $3/$15 rate and $0.30 cached-input rate.
Kimi's model guide says the K2 series will become unavailable to newly registered users on 31 October and is planned for retirement by the end of December, recommending migration to K3.1 Pro. That family-level guidance does not uniquely map TokenScale's two K2.6 tier roles or say K3 is replaced. TokenScale keeps both K2.6 rows visibly carried and K3 freshly verified while Will owns the lineup choices.
Alibaba Model Studio now lists untracked GLM-5.3-Prime at $2.80 input and $8.80 output per million tokens. OpenAI's Decisions API is generally available with GPT-6 Luna, Anthropic-managed Claude Models and Opus 5.5 Fast mode reached Google Cloud, Mistral hosts GLM 5.3 Flash at $0.15 input, $0.03 cached input and $0.50 output, and MiniMax introduced its Code Mini App.
The independent close added four current host-catalog facts: Together now lists MiniMax M3.1 Flash at $0.15 input, $0.015 cached input and $0.60 output; SambaNova now lists it at $0.30 input, $0.03 cached input and $1.20 output; and DeepInfra's current featured catalog now includes GLM-5.3 Flash and DeepSeek V4 Flash. These are API-platform, hosting, catalog or product-surface facts. None uniquely replaces a tracked TokenScale slot, so no Standard price, tracked model, Batch rule or Cache field changed.
OpenAI now offers tracked GPT-5.6-and-later routes an optional 12-hour minimum cache lifetime in the Responses API at a 2.5× input-token write rate. The default 30-minute lifetime remains 1.25× and cache reads remain 0.1× for TokenScale's three tracked OpenAI models. TokenScale records both exact choices without changing the default calculator result or any Standard price.
The same source-watch cluster records Anthropic's 21 October API version and platform-control changes, Cohere's Rerank 5 and Context Focus additions, and Fireworks' Muse Spark Open catalog addition. None uniquely replaces a tracked general-purpose slot, so no lineup or dated Standard point changed.
Two byte-identical fresh reads of Kimi's current Standard and Batch documentation still name Kimi K2.6 and Kimi K3, but both price tables contain empty rows. The Batch page still names K2.6 as supported and says Batch costs 60% of Standard, yet the exact Standard and Batch amounts are absent.
TokenScale carries the three Standard rows at their last verified values, changes the two K2.6 Batch records from supported to unestablished, and treats all three observed Kimi Cache amounts as unestablished. No prior date is rewritten, no zero price is inferred and no replacement model is chosen.
Mistral's official lifecycle table lists Voxtral Mini Transcribe 3, Voxtral Small Transcribe 3, Voxtral Mini Transcribe Realtime 3 and Voxtral TTS 3 as active with a 20 October release date.
These speech transcription, realtime and text-to-speech models are a specialist family, not successors to TokenScale's tracked Mistral Small 4, Medium 3.5 or Large 3 text rows. TokenScale records the Class A launch without changing a tracked price or model.
DeepSeek says DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The same deepseek-flash API alias serves the new checkpoint, and the provider says both the calling method and pricing are unchanged.
TokenScale records the first-party checkpoint identity prospectively without rewriting the earlier V4.1 Flash price history or treating an unchanged rate as a price move. The public lineup keeps its generic V4.1 Flash label pending Will’s naming preference.
Anthropic’s Claude Opus 5.5 Fast mode left research preview and added Microsoft Foundry support. OpenAI’s GPT-Live models gained native Decisions integration, and OpenAI added prompt-to-voice plus nine natural voices. These are processing, integration and realtime-voice changes—not tracked Standard, Batch or Cache repricing.
DeepSeek added Search and Fetch at $0.75 per 1,000 successful requests, with server-side calls through DeepSeek Harness, Claude Code and Codex free until billing begins at 00:00 Beijing time on 1 November 2026. Together’s current catalog lists untracked Reflection Beam 1.0 501B A23B at $0.40 input, $0.04 cached input and $1.20 output per million tokens. Neither ancillary search services nor the untracked model changes a tracked text-price row.
Perplexity’s Agent API catalog added gemini-4-argon, while Cohere’s current catalog changed its North Small Translate identifier to north-small-translate-09-2026. These facts do not resolve Perplexity’s three-slot Sonar basis or replace Cohere’s tracked Command models.
The two held Class B items and five fresh observations crossed the mechanical three-item queue threshold. TokenScale records one factual cluster without inferring a tracked successor, price move or launch date.
Mistral launched Timestral, its first foundation-model family for time-series forecasting. Four models span 5M, 17M, 66M and 209M parameters and are designed to run efficiently on CPUs for zero-shot forecasting.
Timestral forecasts numeric sequences; it is not a successor to TokenScale’s tracked general-purpose text models Mistral Small 4, Medium 3.5 or Large 3. TokenScale records the Class A family launch without inventing a text-model price or changing dated history.
OpenAI scheduled four legacy API families—GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini—to shut down on 14 January 2027 and recommends GPT-6 Luna. None is a current TokenScale tracked OpenAI row, so no dated Standard, Batch or Cache record changed.
xAI launched Grok Voice 3.0 at the same $0.08 per minute plus $0.004 text-input price as Voice 2.0, moved the grok-voice-latest alias and deprecated Voice 2.0 with an explicit migration. It is a specialist voice route, not a successor to a tracked text model.
Together’s current serverless catalog also returned untracked Kimi K3 to $3 input, $0.30 cached input and $15 output per million tokens after the previous close showed $2.70/$0.30/$13.50. The same catalog changed untracked GLM-5.3 Flash quantization from FP8 to FP4 without changing its $0.15/$0.03 cached/$0.50 price. TokenScale records those host facts without projecting them onto Together’s tracked GPT-OSS 120B, Llama 3.3 70B or GLM-5.3 rows.
The four Class B items reached the queue threshold together. They are one clustered source-watch update, not four invented price moves or lineup changes.
OpenAI added generally available Ultrafast processing for GPT-6.1 Sol in the Responses API. It is a processing tier for the same untracked model, not a new Standard-model successor or a change to TokenScale’s Standard, Batch or Cache rows.
Alibaba’s international pricing page now lists the untracked speech model qwen-audio-3.1-tts-flash at $0.23 input and $1.87 output per million tokens. Meta’s current Model API pricing page now lists the untracked Muse Wonder 1.0 at $4 input, $0.20 cached input and $20 output per million tokens. The provider pages do not direct either model into one of TokenScale’s three general-purpose text slots, so no tracked label or dated price point changed.
The oldest item completed its four-night watch and these facts are recorded together. TokenScale has not inferred a launch date where the source only establishes the current catalog.
TypeSafe AI’s Jev is a specialist decision model. It evaluates supplied state against bounded questions and returns typed choices, scores or probabilities. It does not write prose or code, so putting its price beside a model that can write an email, report or novel would make a very cheap number look like a fair comparison when it is not.
TokenScale has recorded the launch without adding a 23rd provider or inventing Small, Mid and Flagship tiers. Vercel AI Gateway currently lists typesafe-ai/jev under promotional free pricing through 25 September. That temporary route is on the dated watchlist for a fresh check when the promotion ends. Check the current Jev model card →
Google replaced its Antigravity managed-agent preview with antigravity-preview-09-2026, deprecated the 05-2026 preview and scheduled the older preview to shut down on 5 October. xAI separately released Grok Voice Transcribe 2.0 at the same $0.10 batch and $0.20 streaming per hour prices as 1.0. Update, 4 October: xAI's 2 October release notes say Transcribe 1.0 reached end of life and all calls now redirect to 2.0 at the same price.
Neither launch replaces a model in TokenScale’s text comparison, and neither changes a tracked text price. They are recorded together after the oldest item completed its four-night watch so useful launch context does not disappear into an internal queue.
Meta's current official pricing catalog lists Muse Spark 1.3, 1.2 and 1.1 on the same Standard basis: $1.25 input, $4.25 output and $0.15 cached input per million tokens. TokenScale still tracks Muse Spark 1.2 in all three comparison tiers. The newer model is recorded here after its four-night watch, but no lineup or historical price was changed because Meta still lists 1.2 and did not direct one unambiguous replacement. Will owns that editorial choice.
Every price on TokenScale is an input cost plus an output cost. Until today the calculator’s single-reply view had no way to say how long the answer was, so it assumed the answer was exactly as long as what you sent. For a text message that is about right. For the larger tasks it was not: picking “The Hobbit” priced the job as though the model reads 95,356 words and then writes a second Hobbit back.
The task presets always carried a realistic answer length — 600 words for a novel, 150 for an email — and the deeper conversation view already used it. The headline number did not. It does now, and the calculator shows that number from the moment you arrive: 1.2K words sent → 300 words back, with the output length adjustable in one tap.
This is a TokenScale modelling correction, not a provider price change. No provider changed anything. Thirteen of the sixteen task presets now read lower — a news article falls by about half, and the long-context tasks by roughly four-fifths. Two presets do not move at all: a text message and a code file, where the answer really is about as long as the request. The pricing labels changed with it, from send and reply to input and output, because half these tasks are not conversations — a photo, a spreadsheet, a transcript — and nothing replies to them.
The 18 August entry correctly recorded DeepSeek V4 Flash at $0.44 input and $1.32 output per million tokens at peak, but TokenScale omitted an important qualifier from its explanation: the 01:00–04:00 and 06:00–10:00 UTC peak windows apply Monday through Friday. Weekends and all other hours are off-peak at $0.22 input and $0.66 output.
This is a TokenScale wording correction, not a provider price change. Against the previous $0.14 input and $0.28 output rate, V4 Flash is about +57% input and +136% output off-peak, or +214% input and +371% output at peak. The dated price observations remain unchanged. Check DeepSeek's official pricing source →
OpenAI lists GPT-5.6 Cyber as an approval-gated model for authorized cybersecurity research and testing, priced at $12.50 input and $75 output per million tokens. It is a specialist, not a general-purpose replacement for GPT-5.6 Sol, Terra or Luna, so TokenScale has recorded the official fact without putting it on the comparison wheel.
This closes a four-night source-watch item. No tracked model changed, no price-history point was created for Cyber, and access still depends on OpenAI approval.
DeepSeek now publishes two standard on-demand windows rather than one flat price. TokenScale uses the highest published rate as the comparison figure and labels it as peak: V4 Flash is $0.44 input and $1.32 output per million tokens; V4 Pro is $1.32/$3.96. The exact off-peak rates are 50% lower outside 01:00–04:00 and 06:00–10:00 UTC. We do not invent an average because each customer's traffic mix is different.
Groq's retired Llama 3.1 8B and Llama 3.3 70B developer rows remain in dated history, but no longer occupy today's board. Small now tracks production GPT-OSS 20B at $0.075/$0.30. Mid and Flagship both track production GPT-OSS 120B at $0.15/$0.60 because Groq currently has only two public production text models. TokenScale does not fill the third tier with a Preview model.
Alibaba still marks Qwen3.7-Max as a limited-time 50% promotion, but its public pricing page does not supply a usable start and end date. TokenScale therefore continues to record the published $2.50 input and $7.50 output list rate, with no backfilled promotion window.
Sakana also added an untracked Namazu model at $0.95 input and $4 output per million tokens. Fugu Ultra remains the tracked flagship at $5/$30. Both facts now sit on the public watchlist for a dated follow-up; neither was turned into an automatic model swap or a made-up price point.
The reliability page exists to count the nights the record missed. From 4 to 9 August it counted them wrong, and wrong in its own favour: it showed 14 carried-forward nights while the price record held 17, and it left the August run marked "ongoing" five days after that run had ended. The nights of 4, 6 and 7 August were never disclosed at all.
What actually happened, from the data. The run that began on 1 August went one night longer than published, through 4 August, and ended on 5 August, when a pass verified 39 of the 66 tracked tiers and held the other 27. A second run followed on 6 and 7 August with no verification pass at all. The 8 August run reached 3 tiers before it stopped. The board has not had a complete nightly pass since 31 July, so that run is left open on the reliability page rather than quietly closed.
The guard matters more than the fix. Counting by hand is what failed, so those counters are no longer typed: they are computed from the price record itself on every build. Each row in the gap table now declares which nights it covers, and the build fails if a night in the record has gone undisclosed for more than three days. The page can still trail the record by up to three nights, and while it does it says so on its face instead of rounding down. The corrected record: 82 nights since 19 May, 63 fully verified, 2 partly verified, 17 carried forward.
First: the live version file, BUILD.txt, could serve a cached copy up to five days stale (it reported v514 while v518 was live). It now ships with a no-store header, so neither browsers nor the CDN may cache it. Second: the nights of 1, 2 and 3 August were never verified. The nightly kept appending data, but no verification pass ran and no provider was stamped after 31 July. Those three nights now stand in the record as carried forward, permanently.
Third, the big one: the methodology page said history is append-only, then said grey carried-forward days "turn solid" once the next check lands. Those can't both be true, and the charts really did repaint held days as verified. The append-only rule wins. Carried-forward nights are now flagged in the data itself, back to May, and the flag never clears: a grey day stays grey, because a later check cannot confirm a night it didn't see. Fourth: the Basic page's no-JavaScript table and its "Rates verified" caption were frozen at 1 July; both are regenerated and stamped 31 July, the true date of the last full verification.
Two additions come with this. A new reliability page lists every night the record missed since 19 May, with the reason: fourteen of seventy-seven, counted honestly. And from 5 August the site's pages are submitted to the Internet Archive nightly by a watchdog running outside the machine the nightly runs on, so there is a third-party, timestamped copy of what this site said that we don't control and can't edit. The full story is in the journal.
Between 13 and 17 July the tracker stopped writing new data, and it took three nights to notice. The cause was mundane: a macOS update reset a permission the background job needs to read its own files, and it failed — quietly — five times in a row. That silence is what we fixed first. The job now raises an alarm the moment it fails, names the actual cause instead of shrugging, and runs on a footing an OS update can’t knock out again. A three-day blackout can’t happen unseen twice.
TokenScale is built to verify every price against the provider’s official page every night. On any night that check can’t complete, the price is carried forward from the last verification so the chart stays whole — and here’s the part worth keeping: those carried-forward days are now drawn in grey, on every chart and table, with each provider stamped at the date it was last actually checked. Every number is real and traces to an official page; nothing was invented, and nothing was lost in the outage. A held price simply shouldn’t wear the confidence of a freshly-checked one — so now it doesn’t.
The methodology page spells this out in full. It’s the same standard we hold providers to: don’t hide the number, and don’t dress an estimate up as a measurement. When we can’t verify a night, we’d rather show you the grey than pretend it’s black and white.
Some models get a launch. Fable 5 got a saga. Anthropic's Mythos-class flagship held our Pro slot for exactly 48 hours in June — 10th to 12th — before a US export-control directive suspended it worldwide, and the board reverted to Opus 4.8. We left those two $10/$50 points in the history because they happened. Yesterday's audit confirmed it: Fable 5 is back on Anthropic's price list at $10/$50, and as of today it retakes the Anthropic Pro slot. Opus 4.8 gets the Sonnet 4.6 treatment — it moves to the all-models view at $5/$25, still listed, still available, nothing deleted.
Read the chart honestly: the Pro line shows the June blip, three weeks of Opus at $5/$25, and a step to $10/$50 today. That step is a model swap, badged as one — the slot changed which model it tracks; nobody's Opus bill doubled. In real terms, the priciest way to write The Hobbit on this board is now about $6.36 (127K output tokens at $50/M), against $0.08 on the cheapest tier. And yes, we note the poetry: the model that now tops the board is the same one that spent yesterday auditing it.
Once in a while the right move is to stop trusting your own machinery and check everything by hand. On 1 July we ran a full drift audit: every one of the 22 providers × 3 tiers on this board, re-verified against the provider's own pricing page, in one sweep, by Claude Fable 5. The score: 40 of 66 tiers checked out exactly. The other 26 are why this entry exists — and in the spirit of this site, we're publishing the misses, not just the fixes.
The worst find is a billing trap, not a typo. Our xAI lite tier still showed Grok 4.1 Fast at $0.20/$0.50. That model was retired on 15 May — but the old API slugs don't error. They silently redirect to Grok 4.3 and bill at $1.25/$2.50, roughly 6× the price you thought you were paying. If you had an agent pointed at that slug, your invoice already knows. The slot now shows Grok Build 0.1 ($1/$2) — xAI no longer sells a sub-dollar text model at all.
Two models on our board never existed. "Kimi K2.6 Turbo" and "MiniMax M2.7 Pro" appear in no official price list — they were plausible-sounding names that slipped in during fast-moving update nights and survived because their prices looked reasonable. A third, Llama 4 Behemoth, was announced in 2025 and never publicly shipped — we were quoting real prices for a ghost. All three are gone: the Kimi slot now tracks K2.7 Code HighSpeed ($1.90/$8), MiniMax gets its actual new flagship M3 ($0.30/$1.20 — flagship capability at lite money, the cheapest flagship tier on the board), and Meta's third slot falls back to Maverick, the biggest Llama you can actually buy.
Then there's the quiet Llama exodus. Eight of our 66 tiers pointed at Llama models their hosts no longer serve: Fireworks has dropped every serverless Llama, Cerebras retired both of its (one the day after our last hand-check), SambaNova's 405B has been gone for over a year, Together delisted the 3B and 405B, DeepInfra dropped the 405B, and Groq's DeepSeek R1 distill was shut down last October. Groq's two remaining Llamas die 16 August — diarised. The GPT-OSS models have quietly become the standard budget tier across the fast-inference hosts, and the board now reflects that. Mistral, meanwhile, had moved a full generation (Small 4, Medium 3.5, Large 3) — and yes, Mistral Large 3 now costs less than Mistral Medium 3.5. Their own FAQ still contradicts their own API table; we've flagged it and gone with the table. Alibaba's whole tracked lineup had been reclassified "legacy, not recommended" by Alibaba itself — replaced with the 3.6/3.7 generation. And Zhipu's "GLM-5-Air" was our mislabel: the $0.20/$1.10 model is GLM-4.5-Air. Right price, wrong name, ours.
The honest bit about our own machinery. The nightly price check did its job — it verified prices. What it couldn't see is that some of those prices belonged to models that had been retired, delisted, or never existed: it was faithfully price-checking ghosts. Three providers' "verified" stamps had quietly frozen at 26 May while the footer said "verified nightly." That's drift, it was ours, and it's the exact failure mode this site exists to catch in others. Fable found it, Fable fixed it, and model-existence checks are joining the nightly routine so the graveyard tends itself. Speaking of which: today's departures — the dead Llamas, the phantom Turbo and Pro, the never-born Behemoth — are all being laid to rest with dates and final prices in the Model Graveyard →
What survived the audit untouched: Google, OpenAI, Anthropic, DeepSeek, Cohere, Perplexity, OpenRouter and Sakana — every tier exact. AWS Bedrock remains the one provider we could not re-verify from source (their pricing page won't render without JavaScript); it's consistent with Anthropic's own global-endpoint pricing, and it keeps its 23 June stamp until we can read it from AWS itself. That's the whole ledger. Next full audit: when the data earns one — the nightly watch continues in the meantime, every night, at Price Moves →
Anthropic launched Claude Sonnet 5 on 30 June. It steps into TokenScale's Anthropic Mid slot in place of Sonnet 4.6, at the same standard list price — $3/M in, $15/M out — and Anthropic says it closes much of the gap to Opus. Sonnet 4.6 isn't deleted: it moves to the "all models" view, still listed and still priced, the way we kept Opus alongside Fable rather than rewriting the record.
Two things worth knowing before you point an agent at it. First, a launch discount: through 31 August, Sonnet 5 runs at $2/M in, $10/M out, then reverts to $3/$15. We track the standing list price as canonical, so the headline number stays $3/$15 and the discount is a temporary window, not history. Second — and very much our beat — Sonnet 5 ships a new tokenizer that turns the same text into roughly 30% more tokens. The per-token price is unchanged, but the same email or novel can cost about a third more to run. That's exactly the kind of hidden shift a sticker price hides and a content-size lens makes visible.
OpenAI unveiled the GPT-5.6 family on 26 June — three models named Sol, Terra and Luna, a new naming scheme that replaces the old mini/nano tiers with capability tiers inside one generation. Access is the story in itself: it's a limited preview reachable only through an OpenAI account rep, after OpenAI shared the models and its release plan with the US government first. General availability is promised "in the coming weeks."
Verified list pricing, per million tokens: Luna $1/$6, Terra $2.50/$15, Sol $5/$30 — Sol matches GPT-5.5's flagship rate, and Terra lands exactly on today's GPT-5.4 mid-tier. Because you can't yet buy these off the shelf, TokenScale keeps GPT-5.5 and GPT-5.4 as the live Lite/Mid/Pro tiers and lists Sol, Terra and Luna in OpenAI's "all models" view, marked preview. We'll promote them to the headline slots the moment they're generally available — the same way we handled Claude Fable 5's brief appearance rather than rewriting the board around a model most people can't reach.
TokenScale now records a small set of anonymous, cookieless engagement events — things like "the page was scrolled past the pricing", "the calculator was used", or "a provider was viewed" — so we can see which parts actually help and build the right things next. In the spirit of the site, we're spelling out precisely what this is and isn't.
What we don't do: no cookies, no IP address stored, no fingerprinting, and the text you paste into the calculator is never sent — it stays in your browser, exactly as before. To follow a single visit's flow we tag its events with a random ID that lives only in your browser's memory, is recreated on every page load, and vanishes when you close the tab; it can't be linked across visits or back to you. We also honour "Do Not Track" and Global Privacy Control — switch either on and we collect nothing at all.
The full detail is on the privacy page →
Zhipu launched GLM-5.2 on 16 June — the first MIT-licensed, 1M-context model to hold a flagship tier. TokenScale's Zhipu GLM flagship slot moves to $1.40/M in, $4.40/M out, against GPT-5.5's $5/$30 at the same context. The +43% over GLM-5.1 is a generational step up, not a repricing.
13 June re-priced five tiers at once, in both directions. Cerebras jumped +299% while AWS Bedrock cut hosted Llama 3.1 405B output from $16 to $2.40 — the night's biggest drop, the opposite direction on the same run. Mistral, Qwen and an OpenRouter route moved too. A clean illustration of how fast the floor shifts.
The 10 June note below recorded Anthropic's Pro slot moving to Fable 5 at $10/$50. That didn't hold. Between 10–12 June the Pro tier round-tripped — $5→$10→$5 in, $25→$50→$25 out — settling back at Opus 4.8 ($5/$25). TokenScale tracks Opus 4.8 as Anthropic's flagship Pro tier; Fable 5 lives in the all-models view, not the headline slot. We're leaving the 10 June entry up as a record of the round-trip rather than rewriting it.
DeepSeek cut its Mid/Pro tier (V4 Pro) by 75% on 11 June — $1.74→$0.435 in, $3.48→$0.87 out. The relentless undercutting that's kept the Novel Index floor near half a cent.
Anthropic launched Claude Fable 5 on 9 June, its first generally available Mythos-class model: a tier above Opus. TokenScale's Anthropic Pro slot now maps to Fable 5 (high) at $10/M in, $50/M out. That makes it the most expensive model on the board, at exactly 2× Opus 4.8. Reading The Hobbit now costs $1.27 on Anthropic's top tier, up from $0.63.
Opus 4.8 isn't gone: it stays in the all-models view alongside Haiku, Sonnet and Fable. Batch pricing halves Fable's rates to $5/$25, and cached reads drop input to $1/M. Worth knowing before you point an agent at a novel.
DeepSeek used to be the only Chinese lab on the board. That stopped making sense. Added four open-weight frontier labs — Alibaba Qwen, Zhipu GLM, Moonshot Kimi and MiniMax — plus Meta's own first-party Llama API, so you no longer have to price Llama through a reseller.
Four of the five are open-weight and price like it — often an order of magnitude below a US frontier flagship. A handful of top-tier prices that aren't public yet are careful estimates, marked for correction. The full story is in the journal →
The month's biggest swings now live on their own page — and each move exports as a branded card you can drop straight into a thread. A percentage tells a developer something; a whole novel going from $1.91 to $0.32 tells everyone.
Each card is drawn on your own device — no server, no tracking, pure static HTML and a little canvas. See all seven moves →
Google updated pricing on their mid-tier Gemini model. Silver (Flash-Lite) stayed the same. Gold (Flash 3.5) did not.
| Model | Input before | Input after | Change |
|---|---|---|---|
| Flash-Lite 2.5 · Silver | $0.10 | $0.10 | — |
| Flash 3.5 · Gold | $0.30 | $1.50 | +400% ↑ |
| Model | Before | After | Difference |
|---|---|---|---|
| Flash-Lite 2.5 · Silver | $0.06 | $0.06 | — |
| Flash 3.5 · Gold | $0.35 | $1.33 | +$0.98 · +380% ↑ |
| Gap between tiers | $0.29 | $1.27 | 4.4× wider |
Hours before posting to Hacker News, we found that our hero number was quoting input cost only. Here's what we fixed — and why it made for a better story.
| What we said | What it actually was | Corrected to |
|---|---|---|
| The Hobbit on Gemini Flash | $0.04 (input only) | $0.06 (total) |
| Input cost | $0.01 | $0.01 |
| Output cost | missing | $0.05 |
| Correct total | $0.04 ✗ | $0.06 ✓ |
The "Which model should I use?" quiz was silently failing on first run. A missing DOM element meant the result screen crashed before anyone could see a recommendation.
The first public version. A single HTML file, no backend, no sign-up. Pricing for 16 AI providers expressed in content you recognise.
| Provider | Model | $/M input | Hobbit cost |
|---|---|---|---|
| DeepSeek | V4 Flash | $0.14 | $0.02 |
| Groq | Llama 3.1 8B | $0.05 | $0.01 |
| Gemini | Flash-Lite 2.5 | $0.10 | $0.06 |
| Mistral | Small | $0.10 | $0.01 |
| Anthropic | Claude Haiku 4.5 | $1.00 | $0.13 |
| OpenAI | GPT-5.4 | $2.50 | $0.32 |
| OpenAI | GPT-5.5 | $5.00 | $0.63 |
| Spread cheapest → most expensive | 35× apart | ||