Four labs, on the record.

Deepest ecosystem; a flagship tied for second at $50 out, a budget model at $0.50

OpenAI's ChatGPT remains the most widely-used AI assistant in the world, and the GPT-6 family is now complete. GPT-6 Astra (preview September 3, wider paid rollout from September 4, 2026) is the flagship at $10/$50 per million tokens, with a catch: requests over 272K input tokens bill the entire request at $20/$75. On the current Artificial Analysis Intelligence Index (v4.3.2, read September 24, 2026) Astra scores 53 at max effort, tied for second with Claude Fable 5.1 and five points behind Claude Opus 5.5 (58), and it ties Fable 5.1 for first on AA's Coding Agent Index at 62. GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50) followed on September 22, 2026, at what OpenAI calls permanent, non-promotional prices, Sol at half of GPT-5.6 Sol's $4/$20 promo rate and Luna at half or more below GPT-5.6 Luna's $0.20/$1.20; AA scores them 48 and 37 at max effort and measured Sol's hallucination rate falling from 92% to 60% against its predecessor. GPT-5.6 Terra is still sold at $2/$12 with no deprecation notice, now undercut on output by GPT-6 Sol. All three GPT-6 models carry a 1,050,000-token context and 128K output. Honest caveats: OpenAI's own system card says Astra "shows a substantial decrease in chain-of-thought monitorability compared to previous models"; it is the first model rated Critical for cybersecurity under OpenAI's Preparedness Framework, so the public version refuses advanced offensive tasks and Enterprise access is off by default; on OpenAI's own charts GPT-6 Sol scores below GPT-5.6 Sol on DeepSWE v1.1 (68.8% vs 72.7%); and Sol and Luna launched in ChatGPT Work and Codex but not in Chat. The ecosystem is the broadest of the four labs: DALL-E, Codex, Atlas browser, app connectors, Memory, Projects, GPT Store, and Microsoft 365 Copilot integration.

First on the Intelligence Index by five points, at $4/$20

Anthropic's Claude Opus 5.5 (September 22, 2026) is the first model in the Claude 5.5 family and the new number one: on the current Artificial Analysis Intelligence Index (v4.3.2, read September 24, 2026) it scores 58 at max effort, five points clear of Claude Fable 5.1 and GPT-6 Astra at 53, with Claude Opus 5 at 51, and it tops GDPval-AA v2.1, AA's real-economic-work eval, at 1846 Elo. It costs $4/$20 per million tokens, 20% less than Opus 5 ($5/$25), with cache reads at $0.20. Anthropic reports 66.4% on Terminal-Bench 4.0 at xhigh effort; Artificial Analysis measured 59.6% on its own run, level with GPT-6 Astra, so we print both and note the gap. Fable 5.1 (September 1, 2026), the safeguarded twin of the trusted-access Claude Mythos 5.1, stays at $10/$50 (cache reads $0.25) and scores 53; Anthropic still tells developers to reach for it on demanding reasoning and long-horizon agentic work when Opus 5.5 at higher effort falls short. Sonnet 5 sits at $2/$10, with Sonnet 5.5 and Haiku 5.5 promised in the coming weeks. All three ship a 1M-token context at standard rates and 128K output. Honest caveats: Anthropic's "40% less than Opus 5" claim is measured at default settings, which means medium effort on Opus 5.5 against high on Opus 5, and at max effort AA saw it use about 119K output tokens per index task against 73K for Opus 5; Opus 5.5 brings five breaking API changes (thinking cannot be disabled, forced tool_choice returns 400, thinking blocks are bound to conversation state for accounts created on or after August 31, 2026, the old computer_20251124 tool is rejected, and inter-tool progress text now lands in thinking blocks); cyber and biology requests fall back to Opus 4.8 or Opus 5 with a visible notice, and API customers must opt in; and Vals AI measured regressions against Opus 5 on legal research, medical coding, and tax work even as Opus 5.5 took #1 on the Vals Index overall.

The agent workhorse, cheapest at its level, half price until December 31

Google's Gemini 3.8 Flash (September 2, 2026) is the third Flash release in six weeks and the current headline model while Gemini 3.5 Pro stays "coming soon" on DeepMind's own site. Artificial Analysis called it the cheapest model at its level of intelligence: on the current Intelligence Index (v4.3.2, read September 24, 2026) it scores 41 at high thinking, two points above 3.7 Flash, at 282.5 tokens/sec output and a $0.58 blended price per million tokens (AA published 59 at launch, but that was the retired pre-September scale). Google's model card shows it beating Claude Opus 5 on 8 of 14 rows (Vals Finance Agent v2 61.4%, Terminal-bench 2.1 89.4%, HLE-Verified 54.9%) and losing the big agentic ones (Terminal-bench 4.0 19.1% vs 51.8%, OSWorld-2.0 (partial score) 59.0% vs 75.4%, GDPVal-AA v2 (pre-re-anchor scale) 1545 vs 1824). Launch pricing is $0.75/$3.75 per million tokens, unchanged from 3.7 Flash, but read the fine print: that is an introductory rate through December 31, 2026, doubling to $1.50/$7.50 on January 1, 2027. 1M-token context, 64K output, multimodal input (text, image, video, audio, PDF), and three thinking levels (low, medium, high). A Gemini 3.8 Flash Cyber variant with looser cyber safeguards is gated to vetted defenders through the Fairwind Program. Honest caveats: time to first token is 15.64 seconds at high thinking per AA, worse than 3.7 Flash; Google says the model uses more tokens per task by design, taking extra reasoning steps and calling tools iteratively, so the per-token price understates the bill; the newest Pro model is still Gemini 3.1 Pro Preview at $2/$12; and Google has announced Gemini 4 pre-training but no date, ID, or price, and has issued no statement cancelling 3.5 Pro despite the rumors. Deep integration across Google Workspace, NotebookLM, Antigravity, AI Mode in Search, and the Gemini app for AI Pro and Ultra subscribers.

A top-four lab at $2/$6, but the per-task token bill is the catch

xAI's Grok 4.7 (September 21, 2026), released under the SpaceXAI name (SpaceX acquired xAI in February 2026, rebranded it SpaceXAI in July, and closed the $60B Cursor deal on August 14), is the lab's most capable model and, per Artificial Analysis, brings SpaceXAI into the top four AI labs: on the current Intelligence Index (v4.3.2, read September 24, 2026) it scores 46 at xhigh effort, #21 of 210 variants, two points above Grok 4.6 (now 44 on the same scale) and behind Claude Opus 5.5 (58), Claude Fable 5.1 and GPT-6 Astra (53), and Claude Opus 5 (51). The stronger story is agentic knowledge work: AA places it #6 on GDPval-AA v2.1 at 1695 Elo and at 1657 on AA-Briefcase v1.1, just behind Opus 5 and Fable 5.1, with an AA-Omniscience hallucination rate of 29% against 51% for GPT-6 Astra. API pricing holds at $2/$6 per million tokens (cached input $0.50), doubling to $4/$12 for the whole request at or above 200K prompt tokens, with a 500K context. It is live in the API, Grok Build (default), Cursor on every plan, and the Grok Office add-ins, but xAI's model card says consumer surfaces (web, mobile, Grok-in-X) come "at a later date". Grok Bot, xAI's always-on AI teammates on a persistent cloud computer, gained Enterprise controls on September 3 and sales connectors on September 10. Honest caveats: AA measured about 81K output tokens per index task, more than double Grok 4.6, so a finished task costs $3.74 against $3.26 for GPT-6 Astra despite the lower per-token price; AA clocks it at 40.4 tokens/sec at xhigh, slower than Grok 4.6, whatever the "same price and speed" launch line says; Terminal-Bench 4.0 is 38.0% at xhigh on xAI's own model card (Grok Build harness) and 33% on AA's run, far off the 59% AA measured for GPT-6 Astra; the 2.1T parameter count is Musk's figure, never published by xAI; and Grok Bot's security model shares one computer and every login across all your Bots (xAI's docs warn against treating separate Bots as a security boundary). Grok 4.3 stays in the lineup as the budget option, and it keeps the 20% batch discount that 4.6 and 4.7 do not get.