Four labs, on the record.

Deepest ecosystem; a flagship tied fourth at $50 out, a near-flagship at $10, a budget model at $0.50

OpenAI's ChatGPT remains the most widely-used AI assistant in the world, and October changed what most of its users get. On the current Artificial Analysis Intelligence Index (v4.3.2, read October 16, 2026) GPT-6 Astra (September 3, 2026) scores 53 at max effort, tied fourth with Claude Fable 5.1 and Gemini 4 Argon and five points behind Claude Opus 5.5 (58), at $10/$50 per million tokens with the whole request billed at $20/$75 above 272K input tokens. GPT-6.1 Sol (September 29, 2026) scores 52 at $2 in / $0.10 cached / $10 out, one point under Astra at $0.72 per index task against $3.26, and it is GA on the API, Azure, Bedrock, GitHub Copilot, ChatGPT Work and Codex but not in the Chat tab. GPT-6 Luna ($0.10/$0.50) scores 38 and now also powers the Decisions API, OpenAI's typed-answer endpoint in public beta since October 6. From October 7 the ChatGPT Chat tab runs GPT-6 Sol with Intelligent UI for paid plans and GPT-6 Luna for Free and Go, one generation behind the Work tab and Codex. DevDay (September 29) added Dots, always-on GPT-6 Astra agents with their own cloud computer, one included on Pro and Business Premium, and a $500 Pro 500 plan with Ultrafast access. Honest caveats: OpenAI did not release GPT-6.1 Astra on September 28 because, in its own words, it "didn't quite meet the bar in terms of staying within scope and authorization", and GPT-6.1 Sol's system card prints worse unwanted persistence (23.5% vs 17.4%) and coding misrepresentation (1.50% vs 0.51%) than Astra; the October ChatGPT builds show statistically significant safety regressions against August's on OpenAI's own benchmarks; extra Dots are unpriced and Pro Dots exclude the EEA, UK, and Switzerland; and the Decisions API ranks 11th of 13 on AA's new Decision Index with no calibration data published. The ecosystem is the broadest of the four labs: Codex, Dots, Atlas browser, app connectors, Memory, Projects, GPT Store, and Microsoft 365 Copilot integration.

First and second on the Intelligence Index, at $4/$20 and $2/$10

Anthropic's Claude 5.5 family is complete: Claude Opus 5.5 (September 22, 2026) leads the current Artificial Analysis Intelligence Index (v4.3.2, read October 16, 2026) at 58 at max effort and tops GDPval-AA v2.1 at 1866 Elo; Claude Sonnet 5.5 (September 28) is second at 56, tied with Meta's Muse Wonder Full, at $2/$10 per million tokens with cache reads cut to $0.10 on October 7; and Claude Haiku 5.5 (October 7) scores 43 at $0.10/$0.50, GPT-6 Luna's list price, with a 1M context. Opus 5.5 costs $4/$20 with cache reads at $0.20, and its fast mode left research preview at $8/$40 on October 15. Fable 5.1 ($10/$50) is now tied fourth at 53 with the highest cost per task of any flagship AA charts. The 5.5 line ships a 1M-token context, 128K output, and a June 2026 cutoff. Honest caveats: at max effort Sonnet 5.5 is the most verbose model AA has measured (420M output tokens to run the index), so its $5.46 per task is within 9% of Opus 5.5's $5.98 and "half the price" is not half the bill; Haiku 5.5 is the only current Claude with a long-context surcharge, 5x to $0.50/$2.50 for the whole request above 100,000 tokens, and it burns about 3x Luna's tokens per task at max; thinking cannot be disabled on Opus 5.5, or above high effort on Sonnet 5.5 and Haiku 5.5, and thinking tokens count toward max_tokens even when hidden; forced tool_choice, assistant prefill, and non-default sampling parameters return 400; Sonnet 5.5 and Opus 5.5 reroute cyber-flagged requests to older models while Haiku 5.5 refuses with no fallback; and the Haiku 5.5 system card reports more hallucination than other recent models and more over-refusal than any model Anthropic tested.

A Gemini 4 flagship at last, tied fourth at $2/$10 until November 30

Google's Gemini 4 Argon (announced September 30 to Fairwind partners, GA October 13, 2026) is the first Gemini 4 model and the first Google flagship since February. On the current Artificial Analysis Intelligence Index (v4.3.2, read October 16, 2026) it scores 53 at high thinking, tied fourth with GPT-6 Astra and Claude Fable 5.1 and five behind Claude Opus 5.5 (58), at $1.99 per index task against Astra's $3.26, and it posts the lowest hallucination rate on AA's board at 15%. It leads the Vals Index at 68.90% and both Zapier AutomationBench boards, and sits #23 on GDPval-AA. Launch pricing is $2 in / $0.10 cached / $10 out per million tokens through November 30, 2026, then $4 / $0.20 / $20 from December 1, with Batch and Flex at half and Priority at 1.8x; the API free tier and the AI Pro plan API do not get it. Gemini 3.8 Flash (September 2) stays the workhorse at 41 on the index and $0.75/$3.75 through December 31, 2026, then $1.50/$7.50. Honest caveats: thinking cannot be turned off on Argon (low, medium, high; default medium), thinking tokens bill as output, and hitting max_output_tokens mid-reasoning returns an empty response that is still billed; the "1M output" headline is a unified 1,048,576-token budget shared by input, thinking, and output, capped at 65,536 on the Gemini API unless you use background mode; AA clocks 44.7 seconds to first token; Google's vendor table runs Argon at highest thinking against vendor-self-reported competitors; and Google confirmed to Reuters that Gemini 3.5 Pro will never ship. Deep integration across Google Workspace, NotebookLM, Antigravity, AI Mode in Search, and the Gemini app for AI Pro and Ultra subscribers, plus the Gemini agent in private preview on Google Cloud, which routes to Claude as well as Gemini.

Fourth-ranked lab at $2/$6, first on voice, and a plan page with no base number

SpaceXSI (the company formerly xAI, renamed SpaceXAI in July and SpaceXSI on October 10, 2026) sells Grok 4.7 (September 21, 2026) as its top text model: on the current Artificial Analysis Intelligence Index (v4.3.2, read October 16, 2026) it scores 46 at xhigh effort, ninth in our table, behind Claude Opus 5.5 (58), Sonnet 5.5 (56), and the trio at 53. Its stronger boards are voice and agentic knowledge work: Grok Voice 3.0 (October 14) is first on AA's Speech to Speech Index at 86.9, ahead of OpenAI's GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6), at an unchanged $0.08 per minute, and Grok 4.7 holds 1711 Elo on GDPval-AA v2.1 with a 29% hallucination rate. 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; Grok 4.7 is on Microsoft Foundry since October 6 and, per third-party reports, reached the consumer Grok app around October 1. On October 12 The X Subscription folded SuperGrok, Cursor, and X Premium into Free, Plus $8, Premium $30, Super $100, and Ultra $200 a month, with Grok Bot on Premium and up. Honest caveats: AA measured 240M output tokens to run the index, so a finished task costs $3.74 against $3.26 for GPT-6 Astra despite the lower per-token price; Grok Voice 3.0 is first on the average but first on only one of its four components, and trails GPT-Live-1 on the support-call benchmark; the subscription's usage multiples (4x, 15x, 50x) have no published base, and Grok Bot has no per-Bot price, Bot count, or named model; Grok Bot's docs still say all your Bots share one cloud computer and "Do not use separate Bots as a security boundary"; Grok 4.8 has a parameter count from Musk and no date; and the 2.1T parameter figure for 4.7 was never published by xAI. Grok 4.3 stays in the lineup as the budget option.