Monthly roundup
October 2026 in AI models
A release-wave summary for readers who want the important changes without pretending every launch is a reason to switch.
The signal
8 releases plus 8 material status updates
Moonshot AI, Xiaomi, Inception, MiniMax, DeepSeek, Google, Meta, Qwen, OpenAI, Anthropic account for 8 tracked release events. 15 catalogued events link to a detailed comparison profile. No dedicated launch guide is published for this month yet. There is no single winner because the constraints are different.
At a glance
- Release events
- 8
- Profile links
- 15
What shipped and changed
Availability updateKimi K3.1 Pro
Moonshot AI
Kimi Code announces full rollout of K3.1 Pro and its Highspeed serving variant.
Date note: This date belongs to Code availability. Platform API IDs, prices and default effort are separate; the API guide promises full weights by November 4, without establishing current download/licence availability. No benchmark placeholders or K3 scores are inherited.
Kimi Code dated release history
Status updateMiMo V2.5 API names expire
Xiaomi
Xiaomi's published schedule expires mimo-v2.5-pro and mimo-v2.5 at 10:00 Beijing time, after an interim replacement-routing period.
Date note: The notice says expired names return errors. Downloaded MIT weights remain separate historical checkpoints. This is the publisher's endpoint schedule, not a live account invocation test or transfer of successor benchmark results.
Xiaomi API replacement and name-expiry notice
ReleaseInception Mercury 3
Inception
Mercury 3 is available on the Inception API with a launch discount and model-specific reasoning defaults.
Inception original Mercury 3 announcement, dated October 19
ReleaseMiniMax M3.1 Flash
MiniMax
MiniMax's dated model log announces M3.1 Flash, the final multimodal model now listed in its API and PAYG documentation.
Date note: The final API ID and downloadable checkpoint are separate from the earlier Preview profile. October 19 is the publisher's announcement date, not an inferred first weight-upload day. Publisher benchmark claims are not imported into independent rankings.
MiniMax dated model release log- Release
MiniMax M3.1 Flash
MiniMax
MiniMax's model release notes announce M3.1 Flash for everyday work.
Date note: The publisher describes open weights. A complete model profile, exact checkpoint licence and pricing still require separate verification; no benchmark result or finder placement is inferred.
MiniMax model release notes 
Status updateDeepSeek V4.1 Flash 1017
DeepSeek
DeepSeek upgraded its rolling deepseek-flash endpoint to the 1017 revision.
Date note: The publisher says API calling and pricing are unchanged. Its new evaluation results use specified vendor harnesses and must not replace earlier checkpoint observations or independent DeepSWE rows.
DeepSeek October 17 update- Status update
MiMo V2.5 interim API routing
Xiaomi
Xiaomi announces temporary automatic routing of V2.5 API names to V2.6 successors with successor billing before the old names expire.
Date note: V2.5-Pro routes to V2.6-Pro; V2.5 routes to V2.6-Flash. The notice conflicts on the routing hour (10:00 in the table, 18:00 in prose), so only the common October 14 date is recorded. October 21 name expiry is a separate event.
Xiaomi API replacement and name-expiry notice 
General availabilityGemini 4 Argon
Google
Google releases the stable gemini-4-argon API endpoint for complex coding, agentic, mathematical and scientific tasks.
Date note: Separate from the September 30 restricted preview. Shared input/thinking/output context is 1,048,576 tokens; synchronous output caps at 65,536, with background long decode up to the remaining context. Introductory paid prices expire November 30; no Free tier. Preview benchmark observations are not a new GA evaluation.
Google Gemini API release notes
ReleaseMuse Wonder and Wonder Full
Meta
Meta's model card dates the release of Muse Wonder and Muse Wonder Full to October 12.
Date note: API Wonder and subscription-only Wonder Full have separate identities and access contracts. This is the card's model-release date; current plan availability is separately checked. No publisher benchmark table is imported into independent rankings.
Meta Muse Wonder model card, page 1
Status updateQwen3-Max QwenCloud retirement
Qwen
QwenCloud's published retirement schedule names qwen3-max and recommends qwen3.7-max.
Date note: The notice gives October 10 at midnight UTC+8, subject to actual rollout timing. This records the announced QwenCloud endpoint retirement, not an all-host shutdown or a live invocation test. Adjacent table limits and prices describe the replacement model.
QwenCloud mainline retirement notice and model-list image- Status update
Gemini 3.5 Flash
Google
Google deprecates the original 3.5 Flash model: its API ID automatically routes requests to Gemini 3.6 Flash and is no longer selectable in AI Studio.
Date note: The alias endpoint is not shutting down. Original checkpoint benchmarks remain separate from successor results and tariffs.
Google Gemini API release notes - Status update
Gemini 3.7 Flash
Google
Google deprecates the original 3.7 Flash model: its API ID automatically routes requests to Gemini 3.8 Flash and is no longer selectable in AI Studio.
Date note: The alias endpoint is not shutting down. Original checkpoint benchmarks remain separate from successor results and tariffs.
Google Gemini API release notes 
Availability updateGPT-6.1 Sol Ultrafast
OpenAI
OpenAI adds Ultrafast processing for GPT-6.1 Sol in the Responses API, with global processing and US/EU data residency.
Date note: A service tier for the existing model, not a new checkpoint. It has separate rate limits and six-times-Standard pricing; measure latency and cost on the actual workload.
OpenAI API changelog
ReleaseClaude Haiku 5.5
Anthropic
Anthropic lists Haiku 5.5 as released October 7, with adaptive thinking and a 1M-token context window.
Date note: Prices depend on prompt length. Standard and beta Batch output limits differ; no Haiku 4.5 benchmark result is inherited.
Anthropic Haiku 5.5 specifications- Status update
GPT-5.3-Codex API deprecation
OpenAI
OpenAI deprecates gpt-5.3-codex, scheduling API removal for April 1, 2027.
Date note: GPT-6 Sol is the recommended replacement. The Codex application's current model menu is a separate product contract.
OpenAI API deprecations - Status update
GPT-5.4 Nano API deprecation
OpenAI
OpenAI deprecates gpt-5.4-nano and its March 17 snapshot, scheduling API removal for April 1, 2027.
Date note: GPT-6 Luna is the recommended replacement. Deprecation is distinct from shutdown; the old model remains separately listed today.
OpenAI API deprecations