Six new models, a cache cut and a shutdown wave: what October 2026 changed on AI bills

October 2026 added six new models to the major model APIs, cut one cache price, and switched off a dozen older OpenAI models on a single day. Mistral put Mistral Large 4 into Public Preview on 6 October. Anthropic launched Claude Haiku 5.5 on 7 October and halved Claude Sonnet 5.5’s cache-read price the same day. Meta launched Muse Wonder on 12 October, and Google made Gemini 4 Argon generally available on 13 October. OpenAI released GPT-6.1 Terra on 26 October and GPT-6 Rosalind, a life-sciences model behind a trusted-access program, on 30 October. In between, on 23 October, OpenAI shut down o4-mini, o1, o3-mini, gpt-4-turbo and the rest of its April deprecation list.

This is the pricing read. It covers what each release lists per million tokens, the fine print that moves the effective rate, which vendors shipped nothing new, and the promotional prices that end in the next eight weeks. Every price below was read from the vendor’s own pricing page or model documentation on 9 November 2026. September’s launches, Claude Fable 5.1, Gemini 3.8 Flash and GPT-6 Astra, are covered in the September edition, linked at the end.

October’s new models, side by side

Three rows on this table match an existing model to the cent. Claude Haiku 5.5 at prompts up to 100,000 tokens lists $0.10 input, $0.50 output and $0.01 cached input, the same as OpenAI’s GPT-6 Luna, released on 22 September. Muse Wonder lists $4.00, $20.00 and $0.20, the same as Claude Opus 5.5. Gemini 4 Argon’s launch price, $2.00, $10.00 and $0.10, is the same as GPT-6.1 Sol and, since 7 October, Claude Sonnet 5.5, until Argon doubles on 1 December. In each case the difference is in the fine print: prompt-length tiers, cache rules, promotion end dates and tokenizers.

Claude Haiku 5.5: a tenth of Haiku 4.5’s list price, with a prompt-length tier

Anthropic released Claude Haiku 5.5 (claude-haiku-5-5) on 7 October 2026 with a 1M-token context window and 128K max output. The release notes list it on the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud (Google managed) and Claude in Microsoft Foundry. For prompts up to 100,000 tokens it lists $0.10 input and $0.50 output, with cache hits at $0.01, 5-minute cache writes at $0.125 and 1-hour writes at $0.20. Claude Haiku 4.5 lists $1.00 and $5.00.

Haiku 5.5 is the only Claude 4.6-or-later model with a prompt-length price tier. A request whose prompt is over 100,000 tokens pays $0.50 input, $2.50 output and $0.05 for cache hits on the whole request. The prompt length counts cache reads and cache writes, so a large cached prefix can push a request over the line. Batch is 50% off either tier.

The tokenizer matters as much as the rate. Anthropic says Haiku 5.5 uses the newer tokenizer from Claude 4.7 onward, so the same text counts as approximately 30 percent more tokens than on Haiku 4.5. Take a workload that measures 10M input and 1M output tokens a day on Haiku 4.5: 10 × $1.00 + 1 × $5.00 = $15.00. On Haiku 5.5 the same text is about 13M and 1.3M tokens. Below the threshold that is 13 × $0.10 + 1.3 × $0.50 = $1.95, 87% lower. If every prompt is over 100,000 tokens it is 13 × $0.50 + 1.3 × $2.50 = $9.75, 35% lower. The 30% is Anthropic’s approximation, and the real increase depends on your content.

Migration also has engineering cost. Anthropic lists changes that return a 400 error on Haiku 5.5 for code written against Haiku 4.5, including manual extended thinking, the temperature, top_p and top_k parameters, and assistant prefill.

Claude Sonnet 5.5: cache reads halved

Also on 7 October, Anthropic lowered Claude Sonnet 5.5’s cache-read price from $0.20 to $0.10 per million tokens, 0.05x the base input price instead of 0.1x. Cache writes and all other prices are unchanged: $2.00 input, $10.00 output and $2.50 for 5-minute cache writes. Claude Opus 5.5 also prices cache hits at 0.05x, or $0.20.

On a retrieval-heavy workload with a 90% cache-hit rate, one million input tokens cost 0.9 × $0.20 + 0.1 × $2.00 = $0.38 before the cut and 0.9 × $0.10 + 0.1 × $2.00 = $0.29 after, about 24% less on the input side. Cache writes are excluded, and a workload that rarely reuses a prefix gets almost none of this.

Gemini 4 Argon: launch price to 30 November, then double

Google made Gemini 4 Argon (gemini-4-argon) generally available on 13 October and calls it its flagship reasoning model. It has a unified 1,048,576-token input and output limit and can generate responses of up to 1M tokens with long decode. The Gemini API price is $2.00 input, $10.00 output including thinking tokens, and $0.10 cached input through 30 November 2026. From 1 December it is $4.00, $20.00 and $0.20, and Batch goes from $1.00 / $5.00 to $2.00 / $10.00.

Three details change the effective rate. Argon is not available on the free tier. It supports implicit caching only, at a 95% discount on cache hits, and Google says explicit caching is not supported. Google also states that every token the model processes is billed, including thinking tokens and the internal tool instructions used for grounding.

On Google Cloud, where Vertex AI is now part of Gemini Enterprise Agent Platform, the pricing page lists the same $2.00 / $10.00 on the global endpoint and $2.20 / $11.00 on non-global endpoints. It says the promotional price is “provided through 50% credits back on net spend”, so on that route the discount may show up as a credit rather than as a lower rate. The same page says that from 13 November, completions over 32K tokens are charged for retrievable token usage even if the request fails on a timeout.

For budgeting, 10M input and 2M output tokens a month cost 10 × $2.00 + 2 × $10.00 = $40.00 in November and 10 × $4.00 + 2 × $20.00 = $80.00 from December, with the same traffic.

GPT-6.1 Terra: half of GPT-5.6 Terra’s input price

OpenAI released GPT-6.1 Terra (gpt-6.1-terra) on 26 October as “a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing.” It lists $1.00 input, $5.00 output and $0.05 cached input, which is 5% of the input rate. Cache writes are $1.25 with 30-minute retention and $2.50 with 12-hour retention. Prompts over 272K input tokens bill the whole request at 2x input and cache rates and 1.5x output, which is $2.00 / $7.50 / $0.10. The context window is 1,050,000 tokens, with a maximum of 922,000 input tokens and 128,000 output tokens. Batch and Flex are 50% lower, and Fast mode is 2x.

GPT-5.6 Terra, which OpenAI names as the replacement for o4-mini, lists $2.00 / $12.00 / $0.20. For 10M input and 2M output tokens, GPT-5.6 Terra costs 10 × $2.00 + 2 × $12.00 = $44.00 and GPT-6.1 Terra costs 10 × $1.00 + 2 × $5.00 = $20.00, 55% less. That assumes the same token counts on both.

The 12-hour write rate comes from an option OpenAI added on 20 October: 12-hour prompt caching in the API, with 12-hour cache writes at 2.5x the input token price. A longer cache only saves money if the cached prefix is read often enough to cover the higher write cost.

Muse Wonder, Mistral Large 4 and GPT-6 Rosalind

Meta launched Muse Wonder on 12 October and said it was “available today” through the Meta Model API and Muse Code. Meta’s documentation says “Muse Wonder is in preview.” The model ID is muse-wonder-1.0, the context window is 1,048,576 tokens, and the price is $4.00 input, $20.00 output and $0.20 cached input. That price is Standard mode only, because Meta’s Contributor mode is not available for Muse Wonder on the Model API. Meta lists no long-context premium.

Mistral’s changelog lists Mistral Large 4 as “available in Public Preview” from 6 October, with a 1M context window and “Launch pricing: 50% off for 2 weeks.” Its model card now shows Status: GA, a release date of 2026-11-05, and $0.70 input, $0.07 cached input and $2.10 output. We have not printed the preview-period price because Mistral no longer publishes it. On 31 October Mistral deprecated Mistral Medium 3.5 and Mistral Large 3, both retiring on 30 April 2027, and named Large 4 as the replacement for each.

GPT-6 Rosalind, released by OpenAI on 30 October, lists $15.00 input, $75.00 output and $1.50 cached input. OpenAI says access is for organizations approved for Rosalind Research through its trusted-access program. The earlier gpt-rosalind-research model, at $5.00 / $25.00 / $0.50, began billing on 5 October.

The 23 October OpenAI shutdowns and other forced migrations

A shutdown is a cost event even when nobody on your team chose it. Any service still calling these models on the day had to move, and the bill now shows whatever it moved to. These are the text-model shutdowns and forced reroutes that took effect in October, with the replacement each vendor names.

Three things about this table. First, OpenAI no longer lists prices for most of the shut-down models, so the before-and-after delta has to come from your own billing history, not the pricing page. Second, three days after o4-mini’s users were pushed toward GPT-5.6 Terra, GPT-6.1 Terra arrived at half its input price and under half its output price, so a migration made in a hurry may already be the expensive choice. Third, Google’s 8 October change was a reroute rather than a shutdown: requests to gemini-3.7-flash and gemini-3.5-flash are automatically routed to 3.8 Flash and 3.6 Flash, which list $0.75 / $3.75 until 31 December and $1.50 / $7.50 from 1 January 2027.

Outside text models, OpenAI also shut down gpt-image-1 on 23 October, naming gpt-image-2.5-sunburst or gpt-image-2.5-flare. Google shut down the Veo 3.1 preview models and gemini-omni-flash-preview on 22 October; its deprecations page now names gemini-omni-1.2-flash, released 5 November, as the replacement, and the deep-research-pro-preview-12-2025 agent on 23 October. Alibaba’s 10 October retirements also covered hosted third-party models, including deepseek-r1, the deepseek-v3 family and glm-4.6 and glm-4.7. It also covered dozens of older Qwen image, audio and snapshot models.

Deprecations announced in October

The biggest announcement for most bills came on 14 October. OpenAI deprecated GPT-4o (gpt-4o-2024-08-06 and gpt-4o-2024-11-20), GPT-4o mini (gpt-4o-mini-2024-07-18), GPT-4.1 (gpt-4.1-2025-04-14) and GPT-4.1 mini (gpt-4.1-mini-2025-04-14). All four are removed from the API on 14 January 2027, and gpt-6-luna is the named replacement for each.

At list price, 10M input and 2M output tokens cost 10 × $2.50 + 2 × $10.00 = $45.00 on gpt-4o and 10 × $0.10 + 2 × $0.50 = $2.00 on gpt-6-luna. A lower list price is not a quality guarantee, and OpenAI does not say the two models produce the same token counts. Run your evaluations before 14 January, not on it.

• ◦OpenAI, 1 October: gpt-5.4-nano, gpt-5.1 and gpt-5.3-codex are deprecated, with removal on 1 April 2027. The named replacements are gpt-6-luna for nano and gpt-6-sol for the other two. tts-1, tts-1-hd and gpt-4o-mini-tts are removed on 6 January 2027, replaced by gpt-realtime-2.1-mini.
• ◦Mistral, 31 October: Mistral Medium 3.5 and Mistral Large 3 are deprecated and retire on 30 April 2027, replaced by Mistral Large 4. Mistral’s API pricing page no longer lists them; their model cards still show $1.50 / $7.50 for Medium 3.5 and $0.50 / $1.50 for Large 3.
• ◦Anthropic, 30 October: fast mode for Claude Opus 4.8 is deprecated and will be removed on 30 November 2026. Anthropic points users to Claude Opus 5.5 for fast mode.
• ◦Google, 6 October: gemini-3.1-flash-image is deprecated with no shutdown date, replaced by Gemini Nano Banana 2.1. Nano Banana 2.1 lists $30.00 per 1M image-output tokens ($0.0336 per 1K image) against $60.00 ($0.067 per 1K image) for 3.1 Flash Image. Its input ($1.50) and text output ($7.50) prices are higher.
• ◦xAI: the October release notes deprecate grok-voice-2.0 in favor of grok-voice-3.0 at the same price.

Who shipped nothing new: xAI, DeepSeek and Alibaba

xAI released no new text model and changed no text-token price in October. Its October model changes were all voice: Grok Voice 3.0 at the same price as Grok Voice 2.0 ($0.08 per minute for speech to speech), and the end of life of grok-voice-transcribe-1.0 on 2 October, with requests routed to version 2.0 at the same price. Grok 4.7, released in September at $2.00 input, $0.50 cached and $6.00 output below 200K prompt tokens, became generally available on Google Cloud’s Agent Platform on 27 October.

DeepSeek released no new model. On 17 October it upgraded the model behind deepseek-flash to DeepSeek-V4.1-Flash-1017 and said “The API calling method and pricing remain unchanged”: $0.15 input, $0.003 cache-hit input and $0.60 output off-peak, doubled at peak hours. The same day it opened its in-house search engine as an API; Search and Fetch are each billed at $0.75 per 1,000 successful requests. Google Cloud’s Model Garden added DeepSeek-V4.1-Flash in Public Preview on 14 October.

Alibaba released no new Qwen text model in October. Its new releases were image and audio models: Qwen-Image-2.1 Pro and Turbo, and the Qwen-Audio-3.1 speech models. Its October text-model entries were existing models reaching new regions. qwen3.8-flash (3 October) and qwen3.8-max and qwen3.7-flash (8 October) became “available with the US deployment scope in the US (Virginia) region”, and qwen3.8-flash was added to Hong Kong and Frankfurt on 27 October. Under the US deployment scope, qwen3.8-max lists $2.00 / $6.00 in US (Virginia), the same as in Singapore; the Global deployment scope in the same region lists $1.65 / $4.951. On 10 October, Qwen3-Max and the other models on Alibaba’s retirement notice ceased service.

Two October changes that show up on the invoice, not the rate card

Anthropic and OpenAI both started giving subscription customers API credits. On 7 October Anthropic said Claude Max, Ultra and Team plans now include monthly API credits. On 28 October OpenAI said eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in free API credits. Credits lower the invoice without lowering usage, so a cost line built from the invoice and one built from tokens × list price will stop agreeing. Track usage cost and credits separately, or a month with credits will look like an efficiency gain.

Billing paths also multiplied. On 22 October Anthropic launched Claude Models on Google Cloud (Anthropic managed), the Claude API reached through a Google Cloud project, where “Google Cloud Marketplace handles billing.” The same Claude usage can now land on an Anthropic invoice, an AWS bill or a Google Cloud bill depending on the route. Separately, on 29 October Anthropic launched a Usage and Cost Analytics API in beta and made the existing Usage and Cost API a legacy API, still supported. If you pull Anthropic costs yourself, plan that migration.

The pattern: promotional prices with end dates

October continued two trends from September. Cached input keeps getting cheaper: Sonnet 5.5 at 0.05x, GPT-6.1 Terra at 5% of input and Argon’s implicit cache at a 95% discount. So the cache-hit rate remains one of the largest cost levers after model choice. The second trend is the one that will catch budgets: several of today’s best prices are promotional and have published end dates. The forecast should use the rate on the date of the spend, not today’s rate.

• ◦13 November 2026: on Google Cloud, Gemini 4 completions over 32K tokens are charged for retrievable token usage even when the request fails on a timeout.
• ◦30 November 2026: OpenAI’s pricing page says GPT-5.6 Sol’s promotional pricing ($4.00 / $20.00) is available until this date. OpenAI does not publish the rate that follows.
• ◦1 December 2026: Gemini 4 Argon goes from $2.00 / $10.00 to $4.00 / $20.00, and cached input from $0.10 to $0.20.
• ◦1 January 2027: Gemini 3.8 Flash and 3.6 Flash go from $0.75 / $3.75 to $1.50 / $7.50. Context-cache storage goes from $0.50 to $1.00 per 1M tokens per hour.
• ◦14 January 2027: gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini snapshots are removed from the OpenAI API.

Tracking what these changes do to your own bill

• ◦Find every caller of a shut-down or rerouted model. A workload silently rerouted from gemini-3.5-flash or pushed off o4-mini is a cost change nobody approved. Per-model spend over time is the view that shows it.
• ◦Measure cache-hit rate per feature before you compare vendors. With cache reads priced between 0.025x and 0.1x of input on the latest Anthropic, OpenAI and Google models, the hit rate can move the bill more than the list-price gap between two vendors.
• ◦Put the promotion end dates in the forecast now. Argon doubles on 1 December, and the Gemini Flash models double on 1 January.
• ◦Reconcile usage against the invoice. Subscription credits, credits-back promotions and marketplace billing all make the invoice differ from tokens × list price.

CloudQuell reads Anthropic usage by workspace, model and token type, and OpenAI usage by project and model, through an admin key. Bedrock and Gemini on Vertex AI appear as AI Spend lenses over the AWS and Google Cloud bills you already connect. Provider data syncs daily and lands on the same ledger as your cloud infrastructure. It is a flat monthly price, and free under $10K/month of tracked spend. To model the changes above before they reach an invoice, use the free LLM pricing explorer, price calendar and cost calculator.

What new AI models were released in October 2026?

Six new models came to the major APIs: Mistral Large 4 in Public Preview on 6 October, Claude Haiku 5.5 on 7 October, Meta’s Muse Wonder on 12 October, Gemini 4 Argon on 13 October, GPT-6.1 Terra on 26 October, and GPT-6 Rosalind, a trusted-access life-sciences model, on 30 October. Anthropic also cut Claude Sonnet 5.5’s cache-read price from $0.20 to $0.10 per million tokens on 7 October.

How much does Claude Haiku 5.5 cost?

For prompts up to 100,000 tokens: $0.10 per million input tokens, $0.50 per million output tokens and $0.01 for cache hits. For prompts over 100,000 tokens the whole request is billed at $0.50 input, $2.50 output and $0.05 for cache hits, and cache reads and writes count toward the prompt length. Batch is 50% off. It uses a tokenizer that produces approximately 30% more tokens for the same text than Claude Haiku 4.5.

When does Gemini 4 Argon pricing go up?

On 1 December 2026. The launch price of $2.00 input, $10.00 output (thinking tokens included) and $0.10 cached input runs through 30 November 2026, then rises to $4.00, $20.00 and $0.20. Batch goes from $1.00 / $5.00 to $2.00 / $10.00 on the same date. Argon is not available on the Gemini API free tier.

Which OpenAI models were shut down on October 23, 2026?

gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-1106-preview, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1, o1-pro, o3-mini, o4-mini and ft-o4-mini-2025-04-16. Fine-tuned versions of these base models follow the same schedule. OpenAI names gpt-5.6-terra for o4-mini and gpt-3.5-turbo, gpt-5.6-luna for gpt-4.1-nano, gpt-5.6-sol for the GPT-4, o1 and o3-mini models, and gpt-image-2.5-sunburst or gpt-image-2.5-flare for gpt-image-1.

How much does GPT-6.1 Terra cost?

$1.00 per million input tokens, $5.00 per million output tokens and $0.05 per million cached input tokens, for prompts up to 272K input tokens. Above 272K input tokens, the whole request bills at $2.00 input, $7.50 output and $0.10 cached. Cache writes are $1.25 for 30-minute retention and $2.50 for 12-hour retention. That is half of GPT-5.6 Terra’s $2.00 input price, and less than half its $12.00 output price.

Did xAI, DeepSeek or Alibaba release new models in October 2026?

Not new text models. xAI’s October model changes were voice-only, with Grok Voice 3.0 at the same price as 2.0. DeepSeek updated deepseek-flash to DeepSeek-V4.1-Flash-1017 on 17 October with pricing unchanged, and added Search and Fetch APIs at $0.75 per 1,000 successful requests each. Alibaba shipped Qwen image and audio models, added existing Qwen3.8 models to more regions, and retired Qwen3-Max on 10 October.