October 2026 AI Model Releases: Claude Haiku 5.5 Leads Price War

A month of cheaper models, not bigger ones

The big AI labs did not use October 2026 to set new capability records. They spent it filling out model families they had launched weeks earlier, making those models cheaper, and widening access to them. The month's biggest frontier release was Anthropic's Claude Haiku 5.5, which arrived on October 732. OpenAI made GPT-6 the default for every ChatGPT user the same day12, and Google put out an image-model refresh while its new flagship stayed behind a restricted program31. In open-weight models, Mistral and Reflection AI both announced frontier-scale systems whose weights have not yet been released2648.

The count varies depending on who is keeping it. One release tracker logged eight new models from eight providers by October 941. Another counted ten models in the first four days alone, all released on October 1 and none from OpenAI, Anthropic, Google, xAI, Meta or DeepSeek43. The gap mostly reflects how each tracker treats speech models, image models and narrow "decision" models, plus the lag between a lab's announcement and a gateway listing43. Every tracker agrees on the basic picture, though. Early October brought a burst of specialist releases, and the big-lab news came from the tail end of September launch cycles.

Anthropic finishes the Claude 5.5 family

The Claude 5.5 rollout came quickly. Opus 5.5 arrived on September 2237, Sonnet 5.5 on September 2831, and Haiku 5.5 on October 7, making it the third 5.5 model in about a month32.

Haiku 5.5 is mainly a pricing move. Anthropic lists it at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens. Above that threshold the rates rise to $0.50 and $2.5033. Anthropic's headline claim is that it costs about 75% less to run than Haiku 4.532. Its footnotes explain where that figure comes from: the price is 90% lower for short requests and 50% lower for long ones, and roughly 90% of Haiku 4.5 traffic fell into the short category34. The model has a one-million-token context window and can produce up to 128,000 tokens of output33. It is also the first Haiku with an adjustable effort setting34.

That price matches GPT-6 Luna, which OpenAI cut to $0.10 and $0.50 on September 2246. One ranking site scores Haiku 5.5 five points above Luna on its composite index46. Anthropic itself still points customers to Sonnet and Opus for complex agentic coding. It pitches Haiku for summarization, compaction and sub-agent work34. Independent leaderboards back up that split: the model ties GLM 5.3 Flash on one Terminal-Bench run but sits only 30th on Arena's WebDev board46.

Anthropic also cut prices elsewhere at the same launch. Cache reads on Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens, and Max and Team subscribers now get monthly API credits34. Several outlets connect this pricing push to Anthropic's planned IPO and to pressure from cheaper Chinese models3339.

The reporting on Opus 5.5 does not fully agree, mostly because "cheaper" can mean different things. Anthropic says token prices fell 20% to $4 input and $20 output, and that typical workloads cost about 40% less because the model uses fewer tokens37. CNBC and Yahoo Finance led with the 40% figure3635. AFP's account, carried by NDTV, framed it as matching Fable at a 20% lower price39. Both figures are accurate. They measure different things.

OpenAI pushes GPT-6 to everyone

OpenAI's October news was about distribution. It did not ship a new frontier model. On October 7 it began putting GPT-6 into the ChatGPT Chat tab: Plus, Pro, Business and Enterprise users get GPT-6 Sol, and Free and Go users get GPT-6 Luna starting October 818. The launch also brought Intelligent UI, which lets answers include charts, forms, buttons and small tools such as calculators18. OpenAI says the instant model starts answering web-search questions 44% sooner than GPT-5.6 Instant did12.

The launches behind this happened in September. GPT-6 Astra arrived on September 315, Sol and Luna followed on September 22 at about half the price of their GPT-5.6 counterparts19, and GPT-6.1 Sol was announced at DevDay on September 2917. OpenAI says 6.1 Sol comes close to Astra on agentic coding and computer use at one-fifth of Astra's standard token price17. It is listed at $2 input and $10 output43. OpenAI also chose not to ship GPT-6.1 Astra. Gizmodo reported the reason as a safety regression1917. A faster tier, GPT-6.1 Sol Ultrafast, reached all API users on October 815.

Google's flagship stays restricted

Google's frontier model, Gemini 4 Argon, was announced on September 30 and is available only to cyber defenders in the company's Fairwind Program71. Google lists an introductory price of $2 input and $10 output per million tokens1. For defenders and its own staff, it says Argon will run without cyber guardrails1. The New York Times called it Google's first significant model since February and an attempt to catch OpenAI and Anthropic, citing Vals AI evaluations in which Argon beat rival models on coding and finance7. TechCrunch noted that Google's comparison benchmarks were Google's own9. CNET described the launch as coming after months of delays4.

In October, all Google actually shipped to the public was Nano Banana 2.1, an image model released on October 6 that replaced Nano Banana 2346. Per-image cost roughly halved, but input-token pricing tripled and the cheap 512px size was removed46. Free Gemini app users also had their default model downgraded from 3.6 Flash to Flash-Lite starting October 946. The strongest model Google offers developers broadly is still Gemini 3.8 Flash, at an introductory $0.75 input and $3.75 output with a one-million-token context10.

Gated releases are now the norm

The three big labs have handled their most capable models in very similar ways. OpenAI put Astra's cyber capabilities behind a trusted-access program13. Google is limiting Argon to vetted defenders1. Anthropic keeps Mythos-class models largely inside Project Glasswing and an expanded, tiered Cyber Verification Program34. Anthropic also says its generally available models block most cyber work by default34.

The timing matters too. Opus 5.5, GPT-6 Sol and GPT-6 Luna were the first releases after Dario Amodei called for an industry-wide slowdown on September 12, a call Sam Altman and Elon Musk joined3639. The labs did not stop shipping. In this analyst's reading, they changed what they ship: cheaper, more efficient versions of models they already had, with the most capable systems held back behind safety gates. That fits both the safety argument and the commercial need to compete with low-cost rivals.

Open-weight models: big announcements, weights still pending

The open-weight side had the month's most ambitious announcements, but neither headline model can be downloaded yet. Mistral Large 4, nicknamed "Le Chonk," is a 1.05-trillion-parameter mixture-of-experts model released on October 6 as a preview API, with weights promised by the end of October26. Its license has not been announced25. The figures for it vary. Reports put active parameters at 49 billion, though Mistral's documentation now says 52 billion46. List pricing of $1.36 input and $4.18 output appears alongside a half-price rate shown on Mistral's own page4644. Mistral claims it is the strongest open-weight model built outside China by a wide margin26.

Reflection AI announced Beam on October 5. It is a text-only model with 501 billion total parameters, 23 billion of them active, and a one-million-token context. The company claims it matches Z.ai's GLM-5.2 while using three to four times less inference compute. Reflection plans to publish the weights this month, and independent verification of its claims is still pending48.

For models that can actually be downloaded today, Chinese labs are still ahead. One October survey ranks GLM-5.3, Kimi K3, DeepSeek V4.1-Flash and Qwen3.8-27B at the top. It names Qwen3.8-27B, released under Apache 2.0, as the most-downloaded new open model of 202624. The same survey cautions that "open" licenses differ: some Qwen and Mistral licenses include revenue or business-type limits2425. Perplexity released a decision model fine-tuned from Qwen3.8-27B under Apache 2.043, another example of Western products being built on Chinese open-weight bases.

Outlook

Trackers expect more big-lab releases in late October and November42. Until then, the market is splitting into two groups. Cheap tiers such as Haiku 5.5, Luna and Gemini Flash now sell for a few cents per million input tokens4610. The frontier models from all three big labs are released in stages and gated. The open-weight test is the end of October. If Mistral and Reflection publish their weights on schedule, Western open-weight models will have a real challenger to the Chinese labs for the first time this year. If either release slips, the month's open-weight news will amount to two API previews.