New AI models in 2026: every major release, tracked

The short answer

As of 23 October 2026, the newest frontier flagships are Google's Gemini 4 Argon (13 October), Anthropic's Claude Opus 5.5 (22 September) and OpenAI's GPT-6 Astra (3 September). If you just want a default: GPT-6 Sol or Claude Sonnet 5.5 for everyday work, GPT-6 Astra, Gemini 4 Argon or Claude Opus 5.5 when correctness matters more than cost, and GPT-6 Luna or Claude Haiku 5.5 when you're paying per token at volume.

The more useful takeaway is what the cluster of releases has in common. In 2026 the frontier stopped being a single leaderboard position and became a price-performance curve: the top models are close enough on capability that cost per task, speed, and tier selection decide which one you should actually run.

What changed in October: ChatGPT's default moved to GPT-6 Sol on paid plans (7 October), and Gemini 4 Argon reached the API and Google's paying subscribers (13 October). Most of the month's other releases were cheaper or more open: Claude Haiku 5.5, DeepSeek V4.1-Flash-1017, MiniMax M3.1 Flash and the Muse Spark Open weights from Meta. Agents also became products in their own right, with parallel agents in GPT-6 Pro and Anthropic's Claude Duet preview.

Figures below reflect each lab's published pricing and reported benchmark results at the time each row was added (the tracker was first written on 26 July 2026). The September and October rows were checked against each vendor's own announcement or documentation, and their prices are as listed at time of writing (October 2026). Model pricing and leaderboard positions in this field change within weeks. Check the vendor's current docs before you commit to one for production work. Last reviewed 23 October 2026.

Every major 2026 release, in order

Two things stand out from that list. First, the release cadence: five frontier launches inside July alone, and that excludes the wider wave of smaller releases (Google's Gemini 3.6 Flash trio, several Qwen models, poolside's Laguna S 2.1) that landed in the same fortnight. Second, the direction: apart from Fable 5, GPT-6 Astra and Gemini 4 Argon, most releases in the second half of 2026 have been about doing more for less, not setting a new capability ceiling. Look at the late-September prices: GPT-6 Sol, GPT-6.1 Sol and Claude Sonnet 5.5 all sit at $2 / $10, and GPT-6 Luna and Claude Haiku 5.5 at $0.10 / $0.50 for shorter prompts.

What actually changed in 2026

Frontier capability got dramatically cheaper

This is the headline. Anthropic positioned Claude Opus 5 as approaching Fable 5's capability in many categories at half the price, keeping the same $5/$25 per million tokens as Opus 4.8. OpenAI made the same move one tier down: GPT-5.6 Terra is pitched as competitive with GPT-5.5 at roughly half the cost, which makes it the obvious migration path for anyone already running 5.5 workloads.

For everyday users this is invisible. You are not paying per token in a chat app. For anyone building with the API, or using AI-powered tools that pass costs through, it is the whole story of 2026.

Every lab ships tiers now

The single flagship model is gone. OpenAI splits GPT-5.6 into Luna (fast, cheap, high volume), Terra (balanced), and Sol (frontier reasoning and long-horizon agent work). Anthropic exposes effort settings so you can trade cost against depth per request, plus a fast mode on Opus 5 that runs about 2.5x faster at double the token price.

The practical consequence: picking the wrong tier now costs more than picking the wrong lab. Running a flagship model to reformat a spreadsheet is the 2026 version of burning money, and running a cheap tier on work that needs careful reasoning is the 2026 version of getting quietly wrong answers.

Coding and agents became the battleground

Grok 4.5 was SpaceXAI's first model built specifically for coding and agentic work, trained using real developer session data, and it reports 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-Bench Pro. OpenAI reported Sol Ultra at 91.9% on the same Terminal-Bench 2.1 benchmark. The benchmarks the labs choose to lead with have shifted from trivia-style knowledge tests toward whether a model can work inside a real codebase over a long session.

Open weights caught up

Moonshot AI published the weights for Kimi K3 on 27 July, roughly 2.8 trillion parameters under a Modified MIT license, the largest open-weight release so far. The practical catch is that running it takes around 594GB of VRAM at full precision, so "open" here means auditable and self-hostable by organisations rather than runnable on your laptop. The pressure it puts on closed-model pricing is real regardless of who can host it.

Benchmarks stopped settling anything

Grok 4.5 landed at 54 on the Artificial Analysis Intelligence Index: fourth place at the time, behind Fable 5, GPT-5.5, and Opus 4.8, at roughly a fifth of the cost of the models above it. Meta's Muse Spark landed fourth on the same index in April. Fourth place at a fifth of the price is a better deal than first place for most real work, which is exactly why the leaderboard alone stopped being a useful buying signal in 2026.

Which model should you actually use?

One caveat worth stating plainly: nobody should reorganize their workflow every time a model ships. If your current tool does the job, a two-point benchmark difference will not change your output. The switching cost is real and the marginal gain usually is not.

What this means if you're learning AI

There is an uncomfortable implication in a list like this: twenty-six major releases in seven months means anything you learn about a specific model has a shelf life measured in weeks. People who tie their AI skills to one product's quirks have to relearn constantly.

What actually transfers between models is the part that has not changed since 2023:

• Prompting structure, giving a model role, context, constraints, and a defined output format works on every model on this page, and will work on the next four.
• Judgment, knowing which tasks a model is reliable for, and which need verification, matters more as models get more confident and more agentic.
• Tool choice, the skill this article is really about: matching the job to the tier, not defaulting to the most expensive option or the one you saw on X.
• Verification habits, the failure mode of a strong model is a plausible wrong answer, and better models make that failure harder to spot, not easier.

That is the case for treating AI as a skill you practice rather than a product you follow. The model names on this page will be outdated by winter. The habits are not.