On October 13, 2026, Google made Gemini 4 Argon generally available, so developers can now choose among three current frontier models: Gemini 4 Argon, Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 family. OpenAI's family includes the flagship GPT-6 Astra and the cheaper GPT-6.1 Sol. This Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6 comparison covers official prices, context and output limits, where each model is available, consumer plan access, and what the published benchmarks show. Vendor and independent benchmarks are kept separate.
This comparison is based on Google's, Anthropic's and OpenAI's official pricing pages, model documentation and launch posts, plus results published by Vals AI, Artificial Analysis and LMArena, all checked on October 14, 2026. For the launch itself, see our Gemini 4 Argon release story.
The short version
- Price: Argon costs $2/$10 per million input/output tokens through November 30, 2026, then $4/$20 from December 1. That standard price matches Claude Opus 5.5. GPT-6.1 Sol costs $2/$10, and GPT-6 Astra costs $10/$50.
- Context: all of them handle about 1M tokens. Argon's limit is combined input and output. Its default output is 65,536 tokens, and up to about 1M with long decode. Opus 5.5, Astra and GPT-6.1 Sol allow up to 128K output tokens.
- Access: Argon is on the Gemini API paid tier (not the free tier), Google Cloud, AI Studio and Antigravity, and in the Gemini app for Google AI Pro and Ultra. Opus 5.5 needs Claude Pro or above. Astra in ChatGPT needs Plus or above.
- Benchmarks: Google's table shows Argon leading on 14 of 19 rows, Anthropic's shows Opus 5.5 leading on Terminal-Bench 4.0, and OpenAI's shows Astra leading on science and computer use. Independent results disagree with each other depending on what they measure.
- Our read: no model wins across the board. Choose by workload, and price Argon at $4/$20, not the introductory rate.
How much do Gemini 4 Argon, Claude Opus 5.5 and GPT-6 cost?
Prices in US dollars per million tokens, standard (synchronous) tier, prompts under each vendor's long-context threshold.
| Model | Input | Cached input | Output | Long prompts | Batch |
|---|
| Gemini 4 Argon, through Nov 30, 2026 | $2.00 | $0.10 | $10.00 | No separate tier listed | $1.00 / $5.00 |
| Gemini 4 Argon, from Dec 1, 2026 | $4.00 | $0.20 | $20.00 | No separate tier listed | $2.00 / $10.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Flat across 1M | $2.00 / $10.00 |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | Flat across 1M | $5.00 / $25.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | Over 272K: $20 / $75 | 50% off |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | Over 272K: $4 / $15 | 50% off |
Last verified: October 14, 2026 (Gemini API, Claude and OpenAI pricing pages; full rows with source links are in our data file M14-frontier-models-compared.csv)
Details that change the bill:
- Argon's introductory price has an end date. Google's pricing page lists $2 input "through November 30, 2026" and $4 "starting December 1, 2026," and every other Argon rate doubles on the same date. Google's September 30 announcement gave no end date, so this is new as of the October 13 launch (our September 30 story covered the gated debut).
- Argon caching is implicit only. The pricing page says explicit caching "is not supported," so you can't pin a long prefix the way you can with Anthropic's cache writes. Google also lists Priority processing at $3.60/$18 now ($7.20/$36 from December 1).
- Thinking is billed as output on all three vendors' models. Google labels its output price "including thinking tokens."
- Tokens per task vary. Artificial Analysis reports that Argon averaged 62,000 output tokens per task on its Intelligence Index, against 27,000 for GPT-6 Astra at max effort. At Argon's standard price, it puts the cost per index task at $3.98, against $3.26 for Astra. Our analysis of why token pricing isn't the story explains why the sticker price misleads.
Worked example (illustrative, not a measurement). One request with a 400,000-token prompt and 8,000 output tokens, at list prices. Google lists no long-prompt surcharge for Argon, so none is applied:
| Model | Cost |
|---|
| GPT-6 Astra (272K surcharge applies) | $8.60 |
| Gemini 4 Argon, standard price | $1.76 |
| Claude Opus 5.5 | $1.76 |
| GPT-6.1 Sol (272K surcharge applies) | $1.72 |
| Gemini 4 Argon, introductory price | $0.88 |
Argon at its standard rate costs the same as Opus 5.5 on this request. The introductory price halves it.
Context windows and output limits
| Model | Context window | Max output | Inputs |
|---|
| Gemini 4 Argon | 1,048,576 tokens, combined input and output | 65,536 by default; up to 1,048,576 with long decode | Text, image, video, audio, PDF |
| Claude Opus 5.5 | 1M tokens | 128,000 | Text, image |
| GPT-6 Astra | 1,050,000 tokens | 128,000 | Text, image |
| GPT-6.1 Sol | 1,050,000 tokens | 128,000 | Text, image |
Last verified: October 14, 2026 (Gemini 4 Argon model page; Anthropic context-window docs; OpenAI model pages)
Google's model page calls Argon's limit a "unified input and output token limit." In practice, a 900,000-token prompt leaves at most about 148,000 tokens for the answer, thinking included, and only with long decode, since the default output cap is 65,536. The 1M-token output only applies to long-decode requests, which have their own documentation.
On knowledge cutoff, Google's model card gives March 2026 for Argon, and OpenAI lists April 30, 2026 for Astra and GPT-6.1 Sol. Anthropic gives a "reliable knowledge cutoff" of June 2026 for its current models, per our Claude lineup guide.
A big window says nothing on its own about how well a model recalls material in the middle of it. Our long-context recall explainer covers why.
Where can you use each model?
| Gemini 4 Argon | Claude Opus 5.5 | GPT-6 Astra / GPT-6.1 Sol |
|---|
| Vendor API | Gemini API, paid tier only; Google AI Studio | Claude API | OpenAI API |
| Clouds | Google Cloud (Gemini for business, Gemini platform API) | AWS, Google Cloud, Microsoft Azure, per Anthropic | Astra: Microsoft Azure, AWS Bedrock |
| Coding tools | Google Antigravity | Claude Code | Codex |
| Restricted variant | Fairwind Program: vetted defenders can get Argon "without cyber guardrails" | Cyber Verification Program for reduced blocking | n/a |
Last verified: October 14, 2026
Google's pricing page says Argon "is not available on the Free tier or on the Google AI Pro plan," which covers the API and AI Studio's subscription access. Google's launch post says Ultra subscribers get Argon in Antigravity and AI Studio, and Pro subscribers get it in the Gemini app. Google is also expanding Fairwind, its unmitigated cyber-access program, to "hundreds of additional partners." DeepMind's program page says it already has more than 650 partners.
Consumer plan access
| Plan | Price (US, monthly) | What you get |
|---|
| Google AI Pro | $19.99 | Argon in the Gemini app and AI Mode |
| Google AI Ultra | $99.99 or $199.99 | Argon with higher limits, plus Antigravity and AI Studio access |
| Claude Free | $0 | No Opus |
| Claude Pro | $20 ($17 billed annually) | Opus 5.5 |
| Claude Max | From $100 | Opus 5.5 with 5x or 20x Pro usage |
| ChatGPT Plus | $20 | Astra, plus GPT-6.1 Sol in Work and Codex; Chat tab runs GPT-6 Sol |
| ChatGPT Pro | $100, $200 or $500 | Astra and "Astra Pro"; Ultrafast only on Pro 500 |
Last verified: October 14, 2026 (Google's subscriptions page and launch post, claude.com/pricing, OpenAI's Pro tiers article and GPT-6 posts)
Free Gemini users don't get Argon. Google's plan pages also show an "Includes Gemini 4 Argon" line at the end of the $4.99 Google AI Plus card, but Plus's own feature list names only Google's Flash model, and Google's launch post and release notes name only Pro and Ultra. We treat Plus access as unconfirmed. For the full ChatGPT plan breakdown, see ChatGPT plans compared.
Benchmarks: what each vendor reports
Every number in this section is vendor-reported. Each company picked the benchmarks, ran its own model and mostly took rivals' scores from those rivals' reports or public leaderboards, under different settings. The tables can't be merged into one ranking.
Google (DeepMind Gemini model page and evaluation methodology, October 2026). Google tests Argon at its highest thinking setting, single attempt. For Opus 5.5, Fable 5.1 and Astra, it uses each vendor's maximum reasoning setting or best reported result. Selected rows:
| Benchmark (what it measures) | Argon | Opus 5.5 | GPT-6 Astra |
|---|
| DeepSWE v1.1 (long-horizon coding) | 77.9% | 74.2% | 74.1% |
| Terminal-bench 4.0 (terminal tasks) | 57.4% | 66.4% | 58.2% |
| FrontierSWE v2 (coding) | 55.0% | 62.3% | 65.5% |
| Terminal-Bench Science 0.1 | 57.6% | 63.3% | 68.1% |
| GraphWalks 256K–1M (long-context reasoning, F1) | 84.2% | 66.8% | 71.8% |
| Harvey's Legal Agent Benchmark | 19.6% | 3.8% | 5.4% |
| CWE-bench v1 (vulnerability fixing) | 68.0% | 67.0% | 68.0% |
Google's own caveats: Argon's OSWorld-2.0 score is the best of three runs, its Terminal-Bench Science run used a verifier timeout six times the default, and video frame budgets differed by model on LVBench. Argon's DeepSWE and Terminal-Bench scores were computed by Google, while rivals' came from public leaderboards or, for the Claude models on DeepSWE, Anthropic's system cards.
Anthropic (Opus 5.5 launch page, September 22, 2026). Anthropic's table doesn't include Argon. It reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 (xhigh effort) against 57.9% for GPT-6 Astra (high effort, as reported by OpenAI). On Terminal-Bench-Science 0.1, it reports 58.7% for Opus 5.5 against 64.6% for Astra. Anthropic ran with production safeguards on, and says fallbacks to older models likely lowered Opus 5.5's scores.
OpenAI (GPT-6 Astra and GPT-6.1 Sol posts, September 2026). OpenAI's tables predate Argon and Opus 5.5. It reports Astra at 72.6% on OSWorld 2.0 (offline, partial) and 64.6% on Terminal-Bench Science, using the best score at any effort level. For GPT-6.1 Sol, OpenAI mostly gives relative gaps rather than scores, such as "Astra minus 2.1 pp" on OSWorld 2.0.
The pattern: on Terminal-Bench 4.0, Google's own table puts Opus 5.5 ahead of Argon (66.4% vs 57.4%), and Artificial Analysis's independent run shows the same order (see below). On long-context and legal tasks, Google's numbers favor Argon by wide margins that no independent party has yet reproduced. Our benchmarks explainer covers what Terminal-Bench and similar suites measure.
What independent evaluators have published
Three evaluators had published Argon results with provenance by October 14, 2026. They measure different things, so they're listed separately.
- Vals AI, Vals Index v2.1 (updated October 13, 2026). This is a GDP-weighted average of agentic tasks in finance, coding, legal and tax. Argon ranks first at 68.9%, then Claude Sonnet 5.5 (67.04%), Opus 5.5 (66.97%), Fable 5.1 (65.83%) and GPT-6 Astra (63.13%). GPT-6.1 Sol is ninth at 61.15%. Vals lists cost per test at $15.68 for Argon (its leaderboard shows Argon's price as $4/$20 but doesn't explain the calculation), $32.14 for Opus 5.5 and $18.46 for Astra. The leaderboard doesn't show each model's reasoning setting. The gaps among the top three are under two points.
- Artificial Analysis, Intelligence Index v4.3.2 (comparison pages checked October 14, 2026; Argon writeup published September 30). Argon (high) scores 53, tied with GPT-6 Astra (max). Claude Opus 5.5, in a configuration Artificial Analysis labels "Max, Default Fallback," scores 58. In its September 30 writeup, Artificial Analysis also reported Argon at 57% on Terminal Bench 4 against 60% for Opus 5.5 (max). It also reported a 15% hallucination rate for Argon on its AA-Omniscience test, against 51% for Astra (max). The index is a composite of the firm's chosen evaluations, and the comparison pages carry no data date.
- LMArena, Text Arena (snapshot dated October 8, 2026, the latest we could confirm). Arena ranks models by crowd votes on paired answers.
gemini-4-argon-high is first at 1525 ±9 on 4,892 votes, a score Arena marks as preliminary. claude-opus-5.5-high is second at 1507 ±8. Voters judge single chat answers, not agentic work, and Argon's sample is still small. On Arena's separate agent board, Opus 5.5 (Max) ranks first and Argon (High) fifth.
We found no published Argon evaluations from METR, Epoch AI, Scale's SEAL leaderboards or the UK AI Security Institute as of October 14.
Which model should you use?
| If you need… | Start with | Why |
|---|
| Lowest frontier-level price until November 30 | Gemini 4 Argon or GPT-6.1 Sol | Both $2/$10; Argon's rises to $4/$20 on December 1 |
| Agentic terminal and coding work | Claude Opus 5.5 | Ahead of Argon on Terminal-Bench 4.0 in Google's table and Artificial Analysis's run, and ahead of Astra in Anthropic's |
| Very long outputs (whole reports, large code migrations) | Gemini 4 Argon | Only model here with up to about 1M output tokens via long decode |
| Long multimodal inputs (video, audio) | Gemini 4 Argon | Accepts video and audio; the others take text and images |
| Hardest science tasks, cost secondary | GPT-6 Astra | Leads Terminal-Bench Science in every table that reports it; $10/$50 |
| Finance, legal and tax agents | Gemini 4 Argon, then Claude | Vals Index ranks Argon first, with Claude models close behind |
| Defensive security work | Apply to Fairwind or Anthropic's Cyber Verification Program | Standard versions block some offensive-security help |
Test on your own tasks before switching. The independent results above disagree with each other, and with the vendors, depending on the workload. The safety trade-offs differ too: see what the Gemini 4 Argon model card says and what the Claude Opus 5.5 system card says. If you want to try Argon, our how-to guide covers setup.
What's confirmed, reported and uncertain
- Confirmed (official pages): all prices and the November 30/December 1 price change, context and output limits, Argon's unified token limit and long decode, Argon's absence from the API free tier, plan prices, and the Pro and Ultra Gemini app access.
- Vendor-reported: every score in the vendor tables, Google's "leads" claims, and Anthropic's and OpenAI's comparisons.
- Independent, with limits: the Vals AI, Artificial Analysis and LMArena results above, each tied to its date and configuration.
- Uncertain:
- whether Google AI Plus subscribers get Argon
- Argon's Gemini app usage limits per plan
- whether Google will keep implicit caching as the only caching option
- results from safety-focused evaluators such as METR and UK AISI, which haven't published on Argon