August 2026 · Notable price changes

Price Moves

August is mostly an audit month. On 16 August, TokenScale corrected two classifications: OpenAI's Luna and Terra reductions were effective 30 July, as OpenAI's dated announcement states, and Sonnet 5 becoming $2/$10 standard was not a customer price cut because users had paid that amount since launch. The earlier entries remain visible as a superseded correction trail. OpenRouter's DeepSeek V3 route did fall again on 12 August. The same 22-route sweep found two directly published launch leads, Grok 4.6 and Kimi K2.7 Code. Alibaba refers to a qwen3.8 family, but no exact public qwen3.8-max-preview row or international rate was reproducible, so TokenScale held every tracked lineup for editorial review. Steps badged model swap are lineup changes, not market moves.

Checked nightly · data through 9 Sep 2026 · Bilton Projects

Overnight in AI pricingdata through 9 Sep 2026 · regenerated nightly from the observation record

No price moves overnight · 17 of 22 providers verified, the rest carried forwardcarried-forward providers are grey · every missed night is listed on /reliability

On the watchlist

OpenAI launched GPT-6 Astra on September 3 at $10 input, $50 output and $1 cached input per million tokens, with a 1.05M-token context window. Rechecked September 9: OpenAI's current model catalog now recommends Astra as its flagship for general use, and the exact model page lists Tier 1 rate limits without the earlier Trusted Access restriction. TokenScale has not changed its tracked lineup; Will owns that editorial model-choice decision.OpenAI official GPT-6 Astra announcement and model page · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Cohere's rendered Generative Models pricing tab reproduced Command R7B at $0.0375/$0.15 and Command R at $0.15/$0.60; the current pricing FAQ reproduced Command R+ (08-2024) at $2.50/$10. The separate current model table still marks all three exact dated SKUs Live, so every tracked rate remains verified.Cohere official pricing and dated model documentation · as of 9 Sep 2026 · until 14 Sep 2026
Rechecked September 9: Google's current Gemini API pricing page says Gemini 3.8 Flash is now available. It publishes a temporary $0.75 input and $3.75 output 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. TokenScale has not changed the tracked lineup; Will owns that model-choice decision.Google Gemini API official pricing and context-caching documentation · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Anthropic's current pricing table continues to publish Claude Fable 5.1 at $10/$50 Standard and $5/$25 Batch, with $0.25 cache read and $12.50/$20 five-minute and one-hour writes. The September 3 guarded successor remains current and prior Fable 5 history stays preserved.Anthropic Claude Platform official release notes, model page and pricing · as of 9 Sep 2026 · until 14 Sep 2026
Rechecked September 9: Together's current pricing table and distinct serverless-model catalog both omit tracked GPT-OSS 20B while listing newer models. TokenScale carries the prior $0.05 input and $0.20 output rate grey; Will owns any replacement-model decision.Together official pricing table and serverless-model catalog · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Fireworks's public model catalog still listed tracked GPT-OSS 20B, but its rendered detail page said Serverless: Not supported and exposed no current rate. The static serverless pricing table also omitted the row. TokenScale carries the prior $0.07 input and $0.30 output rate grey; Will owns any replacement-model decision.Fireworks official serverless pricing table and rendered GPT-OSS 20B model page · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 6: OpenAI's rendered GPT-5.6 Sol page still shows a clearly labelled promotional rate of $4 input and $20 output per million tokens, available at least through November 21. TokenScale keeps the permanent $5/$30 list series unchanged and applies the temporary rate only in the promo-aware view.OpenAI official GPT-5.6 Sol model page · as of 6 Sep 2026 · until 21 Nov 2026
Rechecked September 9: xAI still officially lists Grok 4.6 at $2 input and $6 output per million tokens. TokenScale still tracks Grok 4.5 at the same standard rate pending Will's lineup decision.xAI official pricing · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Moonshot's current Model List still says tracked Kimi K2.5 was officially retired on August 31 and calls now return HTTP 404. It names Kimi K3 or other latest models rather than one unambiguous same-tier successor; K2.6, K2.7 Code and K3 remain current. TokenScale carries the prior K2.5 $0.60/$3 rate grey, and Will owns the Lite replacement decision.Moonshot official Kimi Model List and Batch pricing · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: QwenCloud's public first-party qwen3.8-max model page publishes $2 input and $6 output per million tokens. Qwen3.7-Max remains tracked pending Will's lineup decision.QwenCloud official model page · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: DeepSeek's official pricing page identifies deepseek-v4-pro as the dated DeepSeek-V4-Pro-0813 checkpoint. TokenScale uses the approved peak rate of $1.32 input and $3.96 output per million tokens and keeps the generic V4 Pro label pending lineup and version review.DeepSeek official API pricing · as of 9 Sep 2026 · until 14 Sep 2026
Rechecked September 9: Google's official Gemini API pricing page lists Gemini 3.7 Flash at a temporary $0.75 input and $3.75 output per million tokens through December 31, with a $1.50/$7.50 standard rate beginning January 1, 2027. The tracked Gemini 3.6 Flash row publishes the same temporary and standard rates. The separate $1.50/$9 row belongs to Gemini 3.5 Flash. The lineup decision remains with Will; the temporary-rate start date is not stated, so no promotion was backfilled.Google Gemini API pricing · as of 9 Sep 2026 · until 10 Sep 2026
At the September 9 first desk, Perplexity's current official pricing page reproduced Sonar $1/$1, Sonar Reasoning Pro $2/$8 and Sonar Pro $3/$15 with request fees separate, but no longer reproduced the earlier September 27 support sentence. That mutable lifecycle wording is withdrawn pending the 10 PM closing read; no tracked model or price changed.Perplexity official pricing and current documentation · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: MiniMax's current pricing page omitted exact M2, while the current model material classified it under Legacy Models and did not establish callable eligibility. The September 4 source reversal remains recorded; TokenScale keeps the prior $0.30/$1.20 rate grey and Will owns any replacement decision.MiniMax official pay-as-you-go pricing and current model catalog · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: OpenRouter's exact APIs publish DeepInfra Llama 3.3 70B at $0.10/$0.32, DeepInfra DeepSeek V3 at $0.24/$0.90 and Novita DeepSeek R1 at $0.70/$2.50. The V3 route also publishes $0.135 cached input. TokenScale verified the Mid point without mixing routes.OpenRouter official model and endpoint APIs · as of 9 Sep 2026 · until 14 Sep 2026
Together's rendered table listed tracked Kimi K2.6 at the August 22 8 PM read, but the independent 10 PM closing desk could not reproduce it. Rechecked September 9: the current pricing and serverless-model surfaces still omit the exact Kimi K2.6 inference row while listing newer Kimi models. Those successors cannot be substituted without Will's lineup decision, so TokenScale carries the prior $1.20/$4.50 rate grey until a stable exact row returns.Together official rendered inference pricing table · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Cerebras's complete public model list still contains only Gemma 4 31B and GPT-OSS 120B, while tracked ZAI-GLM-4.7 remains absent. Its exact public endpoint remains unavailable. TokenScale carries the prior $2.25 input and $2.75 output rate grey until Will chooses a current flagship replacement.Cerebras official public model API · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Qwen3.7-Max is officially listed at $2.50 input and $7.50 output per million tokens with a limited-time 50% discount, but the public page does not pin usable start and end dates. TokenScale continues to record the list rate and has not invented a promotion window.Alibaba Cloud Model Studio · as of 9 Sep 2026 · until 10 Sep 2026
Rechecked September 9: Sakana's official pricing page lists the untracked Namazu model at $0.95 input and $4 output per million tokens. Fugu Ultra remains TokenScale's tracked flagship at $5/$30; no automatic lineup change was made.Sakana AI official pricing · as of 9 Sep 2026 · until 14 Sep 2026
Rechecked September 9: Z.ai's official pay-as-you-go table publishes GLM-5.3 at $1.40 input and $4.40 output per million tokens, matching the tracked GLM-5.2 rate. TokenScale continues to track GLM-5.2 until Will makes the lineup decision; the former public-price source gap is resolved.Z.ai official pay-as-you-go pricing · as of 9 Sep 2026 · until 10 Sep 2026

Facts with dates, straight from provider announcements. TokenScale does not speculate.

Back to your view

Compare prices today →

Aug 16TokenScale correctionSonnet 5 standard-price classificationcorrection

Verified Aug 16

$2.00 intro→$2.00 standard in · $10.00 intro→$10.00 standard out

🧙Writing The Hobbit$1.27→$1.27

• Anthropic's official pricing record says Sonnet 5 users paid $2 input and $10 output per million tokens from launch. Making that same amount standard and cancelling a planned future increase did not lower a customer's bill on 12 August. TokenScale previously badged it as a 33% cut at Anthropic and AWS Bedrock. Those two move cards are withdrawn; the $2/$10 current rate remains correct.