Daily Model Brief, October 21, 2026

Mistral ended its sale on Mistral Large 4, so its list price of $1.36 input, $0.14 cached and $4.18 output per million tokens now applies, twice Tuesday's sale price. No other tracked token price changed, and Xiaomi retired MiMo-V2.5-Pro and MiMo-V2.5 overnight.

Wednesday, October 21, 2026 · Prices as of 8:50 AM CT

Executive Summary

What This Means for Customers

• If you call mistral-large-4, budget at $1.36 input, $0.14 cached and $4.18 output per million tokens, twice the sale price that applied through Tuesday. Mistral Large 3 (mistral-large) is unchanged at $0.50 input and $1.50 output.
• If you still call gemini-omni-flash-preview or a Veo 3.1 preview model (veo-3.1-generate-preview, veo-3.1-fast-generate-preview or veo-3.1-lite-generate-preview), finish moving to gemini-omni-1.1-flash today. Google's shutdown is tomorrow, October 22, and Google gives no time of day.
• If you call gpt-4.1-nano, o4-mini or gpt-image-1, they stop at 7:00 PM CT tomorrow, Thursday, October 22. Google's deep-research-pro preview ends October 23, and Perplexity retires its routes to older OpenAI models at 7:00 PM CT on October 23.
• If you called mimo-v2.5-pro or mimo-v2.5, those IDs stopped working at 9:00 PM CT last night. Switch to mimo-v2.6-pro or mimo-v2.6-flash.
• If you reuse long shared prompts on OpenAI's Responses API, 12-hour caching can keep them warm between requests hours apart, but each 12-hour cache write costs 2.5 times the input price. Check that your cache-hit rate covers the write cost before switching.
• If you run non-urgent jobs on grok-4.7, grok-4.6 or grok-4.5, xAI's Batch API pricing page lists a 20% discount for those models on all token types; xAI says most batches complete within 24 hours.
• If your agents search the web through DeepSeek Harness, Claude Code or Codex, those server-side calls start billing at $0.75 per 1,000 successful Search or Fetch requests at 11:00 AM CT on Saturday, October 31. Track this separately from token spend.
• If you run Phi-4 models on Azure Foundry, plan a move before January 17, 2027.
• Claude Haiku 4.5 is still active. Anthropic lists retirement not sooner than October 15, 2026, a date that has passed, and still sets no date. Keep a tested fallback ready.

• One token price changed today: Mistral Large 4. Every price above was checked against the vendor's own page this morning.
• OpenAI's changelog also lists, on October 16, an Ultrafast mode for the Agents API (set agent.service_tier to ultrafast on a supported model) and Artifact SDK skills in OpenAI-hosted sandboxes for spreadsheets, presentations and Word documents. Those entries give no price.
• Anthropic's October 20 release notes: user profiles (new /v1/user_profiles endpoints) let platforms, gateways and resellers tell Anthropic which end customer each request is for, so a safety enforcement action can apply to one end customer instead of the whole organization. Organization webhooks are in beta on the Admin API, and inference hooks are out of beta for Claude Enterprise. None carries a price.
• Cohere added Rerank 5, in Pro and Fast tiers, with a Context Focus feature that returns the most relevant passages within each document. Cohere's models page shows no price for it.
• Google's pricing, models and deprecation pages are unchanged since Tuesday, including Gemini 4 Argon's launch price through November 30 and Grounding with Google Search at 5,000 free requests per month, then $14 per 1,000.
• Amazon's Bedrock price pages could not be read today (they land on an overview page with no prices), ByteDance's price page shows navigation only, Cohere still publishes no full token matrix, Microsoft's Foundry catalog lists no pay-as-you-go prices, and IBM's supported-models docs returned 403. Their last confirmed prices stand.

Deprecations

• Perplexity will retire its Agent API and Orchestrator API routes to older OpenAI models (openai/gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1, gpt-5 and gpt-5-mini) at 7:00 PM CT on Fri Oct 23, 2026 (00:00 UTC Oct 24). Unchanged.
• Microsoft's Foundry retirement page still lists claude-opus-5-5-fast (version 2) on Azure retiring Fri Jan 15, 2027, with no replacement named.
• OpenAI's earlier notices are unchanged: gpt-5.3-codex, gpt-5.1 and gpt-5.4-nano shut down April 1, 2027, and the /v1/audio/translations endpoint shuts down February 26, 2027, replaced by v1/realtime/translations.
• No other shutdown dates moved since Tuesday at Anthropic, Google, OpenAI or Cohere.
• Unverified: Google's deprecations page still shows gemini-2.5-flash-image shutting down Mon Mar 15, 2027, with gemini-3.1-flash-lite-image as the replacement, while Google's pricing page says it shut down Oct 2, 2026. Confirm with Google before you budget either date.

Agent Platforms

Hosted AI agents that come with their own persistent cloud computer. Each line comes from the maker's own pages, read this morning.

• Grok Bot: its own persistent cloud computer, one per user, with no separate computer charge found; usage comes from a weekly pool, then prepaid top-ups. No change since its October 14 release.
• dots (OpenAI): OpenAI says a dot "has its own cloud computer"; the first dot comes with Pro and Business Premium "at no extra cost." Runs on GPT-6 Astra. OpenAI's help page was updated this week; texting is a beta limited to US Pro users.
• Muse (Meta): "You and your Muse share your own dedicated computer in the cloud." Runs on Muse Spark 1.3, whose API price is unchanged at $1.25 input, $0.15 cached and $4.25 output per million tokens. Billing is unverified today; Meta's subscription page did not load.
• Manus: ordinary tasks get a temporary sandbox; the Cloud Computer add-on is a persistent server billed separately at $30 a month (2 vCPU, 4 GB) or $50 a month (2 vCPU, 8 GB). No change.
• Genspark Claw: a "dedicated cloud computer provisioned just for you," billed separately as a Lite, Standard or Powerful Cloud Computer subscription plus credits; Genspark's help page gives no dollar amounts. No change.

Model Size vs. Price

Across the 20 active models whose vendors publish parameter counts, price rises only loosely with total size (rank correlation 0.41 on output price). Anthropic, OpenAI, xAI and Google publish no parameter counts, so their models, including Claude Haiku 5.5, GPT-6.1 Sol and Gemini 4 Argon, are not plotted. The one plotted price that changed since Tuesday is Mistral Large 4, now at its $1.36 / $4.18 list price. Prices as of October 21, 2026.

Price vs. Performance

Across the 45 active models with a measured Artificial Analysis Intelligence Index score, price rises with performance only loosely (rank correlation 0.46 on output price). Claude Haiku 5.5 leads at $0.50 or less (43.4); GPT-6.1 Sol scores 51.8 and Gemini 4 Argon 52.6, each at $10. Scores as fetched October 16, 2026, 8:47 AM CT, the most recent complete leaderboard read; prices as of October 21, 2026.

All times Central. Vendor prices, dates and model IDs are taken from first-party vendor pricing, model, deprecation and changelog pages as of 8:50 AM CT on October 21, 2026. Chart prices are as of October 21, 2026. Performance scores are from the Artificial Analysis Intelligence Index, fetched 8:47 AM CT on October 16, 2026 (the most recent complete leaderboard read). Agent platform details come from each maker's own help, blog and pricing pages. Items that could not be confirmed are marked unverified.