New AI Models in October 2026: Gemini 4 Argon Opens, Prices Fall

Small models led October, then the big labs followed

October 2026 began with a run of narrow, specialized models instead of a big flagship launch. The large labs only stepped in later. By the second week, Google had opened Gemini 4 Argon to paying developers, Anthropic had finished its 5.5 family with a much cheaper Haiku, OpenAI had made GPT-6 the default in ChatGPT, and Mistral had previewed a trillion-parameter model. The most important change this month was not any single benchmark score. It was how far the price of capable models has fallen, along with a growing gap between the day a model is announced and the day ordinary customers can use it.

The trackers do not agree on how many models have shipped. One gateway counted 12 releases from 12 providers by October 122. Another ledger counted ten as of October 4, all released on October 1, and noted that no OpenAI, Anthropic, Google, xAI, Meta or DeepSeek model had launched in October up to that point8. The gap comes mostly from definitions: whether a gateway listing counts as a release, and whether image, voice and video models count as "models." One tracker chose to log three OpenRouter entries that had no vendor release page as listings rather than releases6. That kind of care matters, because headline counts are now padded with relistings and research previews.

The October 1 wave of specialist models

The first day of the month was full of task-specific systems. Cloudflare released Clef and Clef-flash, which it calls decision models. They return typed answers with probabilities instead of free text, come in 27B and 9B sizes, and were post-trained from Qwen bases6. Amazon's Strands Labs released Strands Decider 2B, a decision model built to run on a local CPU or GPU, and published the weights, training data and scripts6. Perplexity's pplx-decider-v1-27b was the cheapest of these, at $0.04 per million input tokens with no charge for output8.

Microsoft AI released MAI-Transcribe-2-Streaming, a streaming speech-to-text model covering 60 languages, along with two text-to-speech models, MAI-Voice-2.1 and a low-latency Flash version6. Prices were $0.54 per audio hour (an introductory rate), $22 per million characters and $15 per million characters respectively8. Black Forest Labs released FLUX 3 Image, and Tavus opened Griffin-Lite, a full-duplex video model, to invited testers only, with no published price8. On October 2, Bilibili's Index team published a preview translation model with 35B total parameters, 3B active, and support for 150 languages6.

Four of these early releases use the Apache 2.0 licence and have public weights3. Taken together, they suggest that decision-making, speech and translation are becoming separate product categories, priced by the task or the character rather than by the token of open-ended chat.

Gemini 4 Argon opens to paying customers

Google's Gemini 4 Argon was the most anticipated release. Google announced it on September 30, but at first only trusted cyber defenders in its Fairwind Program could use it, and Google said API access would come "soon"13. On October 13, Reuters reported that Google had released Argon to developers and businesses worldwide. Google described the model as built for "deep work" on complex, multi-step tasks in coding, finance, law, cybersecurity and science15. Paying customers can reach it through Google's app-building tools. Google AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search, while the Antigravity coding tool is limited to Ultra subscribers15.

The context is a difficult year for Google. In September, Google dropped plans to ship Gemini 3.5 Pro and reorganized DeepMind after Demis Hassabis stepped aside and several Gemini leaders left12. Before Argon, Google's newest public model was Gemini 3.8 Flash, which one analysis said trailed GPT-6 Astra and Claude Opus 5.5 on independent intelligence benchmarks16. The same analysis cited CNBC reporting that Argon tied OpenAI on a cybersecurity benchmark and led on software-engineering results16. Those comparisons have not yet been settled by independent testing.

Price is the clearest part of Google's pitch. Argon starts at $2 per million input tokens and $10 per million output tokens, and Reuters reported those rates will double after November 3015. List pricing rises to $4 and $2016. Times Now described the introductory rate as roughly one-fifth of what GPT-6 Astra and Anthropic's Fable models cost20. That matches published price sheets, which list GPT-6 Astra at $10/$5022 and Claude Fable 5.1 at $10/$5027. Even after the increase, Argon would sit at the same price as Claude Opus 5.5 at $4/$2027. Any team budgeting around Argon today should assume the higher rate.

Anthropic finishes the 5.5 family with a cheap Haiku

Anthropic released Claude Haiku 5.5 on October 7. Trackers describe it as the last member of the 5.5 family, aimed at high-volume work such as summaries, classification and live customer support, and the first Haiku with an adjustable effort setting9. It is listed at $0.10 input and $0.50 output per million tokens with a 1-million-token context window4. That is far cheaper than Haiku 4.5's $1/$527. One rate card notes that the $0.10/$0.50 price applies up to 100K tokens, and the price rises to $0.50/$2.50 above that13. Builders using long contexts should read that detail closely.

This is the most direct competitive move of the month. Haiku 5.5 matches GPT-6 Luna exactly on list price24. A month earlier, one comparison had calculated that Luna was about ten times cheaper than Haiku 4.5 on the workloads it modelled. Anthropic has closed that gap completely. Separately, Anthropic confirmed monthly API credits for Claude Max and Team subscribers21.

Reports disagree on OpenAI's "price cut"

OpenAI started rolling out GPT-6 with what it calls Intelligent UI on October 7. Answers can mix text with charts, forms and interactive elements. Paid ChatGPT tiers use GPT-6 Sol, and Free and Go users, who got access a day later, use GPT-6 Luna79.

Coverage of the API side conflicts. One outlet framed the rollout as OpenAI halving its API prices, with GPT-6 Sol at $2/$10 compared with GPT-5.6 Sol's $4/$2023. A pricing tracker reported that the October 7 rollout came with no new subscription or API rate21. Another rate card dated GPT-6 Sol and Luna pricing to their September 22 launch13. The more accurate reading is that the 50% cut is a comparison between generations that took effect in September, not a new price move in October. GPT-5.6 Sol remains on the price sheet at a promotional $4/$20 through at least November 2122.

OpenAI's actual October price changes were smaller. GPT-6.1 Sol replaced GPT-6 Sol on the price list on October 4. It keeps the same $2/$10 standard rates but cuts cached input to $0.10 per million tokens25. A research endpoint, gpt-rosalind-research, also began billing on October 5 at $5 per million input tokens, limited to approved researchers.

Open weights, mostly promised for later

Several open-weight releases were announced this month, but most of the weights have not shipped yet. Mistral previewed Large 4 on October 6 and said the weights would follow by the end of the month7. Reports differ on its design: one describes 1 trillion total parameters with 49B active3, and another gives roughly 52B active7. Reports also differ on price. One account said no price had been published at preview3, while rate cards now list $0.68 input and $2.09 output with a 1M context window426.

Reflection AI, a Brooklyn startup, unveiled Beam on October 5. It is a 501-billion-parameter mixture-of-experts model with 23B active parameters. Reflection says it matches Z.ai's GLM-5.2 on reasoning while using three to four times less inference compute, but those claims have not been independently verified, and the weights are also promised for later in the month10. One release that did ship is POCKET-Darwin-180B, a four-bit build of a 180B model published October 2 that its developers say can run from a laptop SSD3. Elsewhere, StepFun opened a Step 5 Preview API at $1.00/$2.704, and Meta released Muse Wonder on October 122.

What it means for buyers

The overall direction is clear. One index puts frontier token prices 84% below their March 2023 baseline24. Anthropic's argument for Opus 5.5 is that it uses fewer tokens per task, so typical workloads cost about 40% less than on Opus 5, even though the list price fell only 20%. In practice, that makes cost per completed task a more useful number than price per token.

The main risk this month is price increases that are already scheduled. Argon doubles after November 3015, GPT-5.6 Sol's promotion ends after November 2122, and Gemini 3.8 Flash rises from $0.75/$3.75 to $1.50/$7.50 on January 1, 202727. More of the big labs' October launches were cheap at the low end of their lineups than at the top, and the competition has moved to the entry tier, where Anthropic and OpenAI now charge the same. For flagship models, the low prices on offer this month are mostly introductory rates with an expiry date.