AI models in September 2026: who leads, what it costs, what to choose

The closed labs and their flagships

Who sells what, since when and at what list price.

Anthropic. Claude Fable 5.1 is its most capable widely released model; its twin, Mythos 5.1, is limited to verified US organizations. Anthropic's docs now recommend starting with Opus 5.5 for most workloads, and Sonnet 5.5 and Haiku 5.5 are announced for the coming weeks (details in our Opus 5.5 analysis).

OpenAI. GPT-6 Astra is its first model rated "Critical" for cybersecurity under its Preparedness Framework, so the public version refuses advanced offensive tasks. GPT-6 Sol and Luna launched at half the promotional price of their GPT-5.6 predecessors; GPT-5.6 Terra ($2/$12) remains the middle option and the o-series is now legacy.

Google. Gemini 3.5 Pro, expected in June, is late: Bloomberg reported on July 16 that its coding performance fell short. Meanwhile a new Flash model shipped almost every month (3.5 in May, 3.6 in July, 3.7 in August, 3.8 in September), and Gemini 3.1 Pro is still in preview.

SpaceXAI. SpaceX absorbed xAI in February 2026, renamed it in July and bought the code editor Cursor in August. Grok 4.7 ranks sixth on the official Terminal-Bench 4.0 board (37.6%).

Meta, Microsoft, Amazon. Muse Spark (April 8) was Meta's first model without open weights; version 1.3 scores 48 on the Artificial Analysis index. Microsoft unveiled MAI-Thinking-1, its first reasoning model, in June (price unconfirmed). Amazon moved Nova Premier and other Nova models to maintenance mode, according to Business Insider (via eWeek).

Open weights in brief

Open-weight models have grown huge (Kimi K3 has 2.8 trillion parameters, 104 billion active per token) but still trail the closed leaders: the best, Xiaomi's MiMo-V2.6-Pro, scores 46 on the Artificial Analysis index against 58 for Claude Opus 5.5.

DeepSeek V4.1-Flash (MIT license) is the cheapest near-frontier API we found, at $0.30/$1.20 at peak hours and half that off-peak, and licenses are drifting toward custom terms with revenue thresholds. More in our open-weight guide.

Price bands: four tiers, not one market

The list price tells you the band; your cost per task tells you the bill.

The $10/$50 tier is a 2026 novelty (Claude Fable 5 in June, then Fable 5.1 and GPT-6 Astra), while the middle got cheaper: GPT-6 Sol costs half its predecessor's promotional price, Opus 5.5 is 20% cheaper per token than Opus 5, and Sonnet 5 kept $2/$10 instead of rising to $3/$15.

Three details move a real bill more than the list price: long prompts (OpenAI doubles the input rate above 272K tokens; Claude has no surcharge), cache reads (5% of the input price on Opus 5.5, 10% on GPT-6) and scheduled rises (Gemini 3.6 to 3.8 Flash double on January 1, 2027). Method in our AI cost guide.

Leaderboards: what they measure, who leads

Each ranking answers a narrower question than "which model is best".

Vendor tables are run or chosen by the lab itself, usually at its highest effort and on benchmark versions a few weeks old. Independent boards apply one method to everyone but lag behind: on September 26, Anthropic's reported 66.4% for Opus 5.5 on Terminal-Bench 4.0 was not yet on the official board.

Versions also break comparisons: OpenAI's Astra post quotes Artificial Analysis index v4.1.1 values of up to 65.7, while v4.3.2 tops out at 58, a new scale rather than weaker models. Arena, for its part, ranks by human preference votes, not by tasks completed.

None of these tests uses business Spanish or French: they filter a shortlist, they don't decide it. Our method is in the evals guide.

How to choose by use case

A shortlist to test, not a verdict.

Include a second vendor and an open-weight option in every shortlist. Our method (30 fixed prompts per client, five engines per run, three languages) is described in how we choose models.

Seven verified trends of 2026

• A premium tier, yet cheaper capability. On ARC-AGI-2, a score of about 61 to 65% cost $2.25 per task with Claude Opus 4.6 in February and $0.042 with DeepSeek V4 Flash in July.
• Reasoning effort is a setting. Claude, GPT-5.6, GPT-6 and Gemini 3.x expose effort levels; on ARC-AGI-2, Opus 5.5 costs $0.241 per task at low effort and $1.85 at max (see our reasoning guide).
• 1M tokens of context is the norm for flagships; Grok 4.7 stops at 500K.
• Offensive cyber capability is gated. Mythos 5.1 is limited to verified US organizations, Gemini 3.8 Flash Cyber to trusted defenders in Google's Fairwind Program, and OpenAI plans wider defensive access to Astra through its Daybreak program.
• Products disappear quickly. OpenAI closed the Sora app in April, about seven months after launch, removed the Sora 2 API on September 24 and shut its Atlas browser on August 9; Nova Premier reached end of life on September 14.
• US policy can cut access. A US export-control directive made Anthropic pull Fable 5 and Mythos 5 for all users on June 12 (Fable 5 returned on July 1), and Executive Order 14409 (June 2) calls for a voluntary framework giving the government up to 30 days of pre-release access to "covered frontier models".
• Labs debate a slower pace. In We Must Pace the Frontier, Dario Amodei called for slowing capability gains and committed Anthropic to embedded third-party evaluators; Sam Altman matched that commitment on September 14. One trigger he cites: the July incident in which OpenAI agents under evaluation reached Hugging Face systems (OpenAI's report).

What it means for companies in Colombia, Latin America and Europe

Keep a tested fallback. The Fable 5 withdrawal lasted 19 days and hit every customer. A production system in Bogotá, Mexico City or Paris needs a second model from another vendor, tested on the same prompts, and an open-weight option you could host.

Budget for residency. OpenAI adds 10% for regional processing on models released since March 5, 2026; Claude costs 10% more on regional Bedrock or Google Cloud endpoints.

Put retirement dates in contracts. Anthropic commits to Opus 5.5 until at least September 22, 2027, but to Haiku 4.5 only until October 15, 2026. Ask every vendor for such dates, notice periods and a migration clause.

Test Spanish and French yourself. At Slash AI Lab every candidate goes through 30 fixed prompts per client in Spanish, English and French before deployment; no leaderboard replaces that step.

Check the reach of the EU AI Act. Article 2(1)(c) covers providers and deployers outside the EU when their system's output is used in the EU, so a Colombian firm serving French clients can be in scope. In Colombia, check how Law 1581 of 2012 applies before sending personal data abroad. More in our 2026 regulatory map.