One best model is the wrong goal

People often ask which AI model is the best. That question is too wide. Research is not one job. Research is a chain of jobs. The chain runs from idea to plan, then test, check, and write-up. Each job needs something different.

So this workflow uses three models, each with a fixed job:

  • GPT-6 Astra, through ChatGPT: ideas and new system designs.
  • Claude Opus 5.5: research planning with a human in the loop, plus running experiments.
  • Gemini 4 Argon: documents and writing.

This post explains why. Public test scores appear where available. Personal preference fills the gaps, and the text says which is which. All facts are from early November 2026. This field moves fast, so check the date before trusting a number.

Which one to choose?

One research project, three kinds of help.

The three models at a glance

GPT-6 Astra Claude Opus 5.5 Gemini 4 Argon
Released 3 Sep 2026 22 Sep 2026 Announced 30 Sep 2026
Developer price per 1M tokens (input / output) $10 / $50 $4 / $20 $4 / $20, launch offer $2 / $10
Average cost per task $3.26 $5.98 $3.98, or $1.99 with the offer
Main job here Ideas, new systems Planning, experiments Documents, writing

Tokens are small pieces of text, and companies charge by the token. The token price is not the real price, though. Artificial Analysis measured about 62,000 output tokens per task for Argon. Astra used about 27,000. So Astra can cost less per task, even with pricier tokens. These tests used the highest effort settings, so normal use costs less. Opus at top effort cost the most of the three.

How the three fit together

The chart below shows the path from a question to finished work.

flowchart TD
    A["Research question"] --> B["GPT-6 Astra: find and test ideas"]
    B --> C{"Is the idea worth testing?"}
    C -->|No| B
    C -->|Yes| D["Claude Opus 5.5: draft the plan"]
    D --> E{"Human review: is the plan good?"}
    E -->|No, change the plan| D
    E -->|Yes| F["Claude Opus 5.5: run the experiments"]
    F --> G{"Human check: do the results make sense?"}
    G -->|No, fix or rethink| D
    G -->|Yes| H["Gemini 4 Argon: write documents and drafts"]
    H --> K["Human check: verify every claim and source"]
    K --> J["Share the work"]

Each diamond is a point where a human stops and decides. No model in this chain gets the final say.

GPT-6 Astra: ideas and new systems

ChatGPT comes first when the goal is still unclear. A good opening request is ten ways to attack a problem. The next request asks Astra to tear apart the favourite idea. The same method works for designing a new system. Ask what the parts are, where the system could break, and what nobody has tried.

Some evidence supports this choice. OpenAI reports 97.6% on FrontierMath Tier 4, a very hard math test. Claude Fable 5.1 scored 87.8% on the same test. OpenAI also shared new results on gaps between prime numbers. No independent check turned up, so those results remain a claim.

Two warnings apply. First, the test setup can change a score. Astra's 99.9% on ARC-AGI-3 used a special setup. Independent runs of the standard setup scored roughly 17% to 63%. Second, no test measures a good idea. Expert testers rated Opus 5.5's new ideas as about equal to the best model tested. That report comes from Anthropic, but Astra clearly has rivals here.

So this choice is mostly preference. Astra offers a wide spread of ideas and pushes back well. Astra also gets more facts wrong than Argon, as shown below. That is fine at this stage. A wrong idea costs little when a person checks the idea next. A wrong fact in a final report costs far more.

Claude Opus 5.5: planning and experiments

Opus handles the long, careful work while a human stays in charge. The routine is simple. Opus writes the plan first: the main question, the data, and the measures. The plan also says what result would prove the idea wrong. A human reads and edits the plan before any code runs. Only then does Opus run the experiments, reporting back at each stage.

The numbers support this choice well:

  • Terminal-Bench 4.0, a test of long computer tasks: Opus scored 66.4% in Anthropic's table. Artificial Analysis measured about 60%. Even Google's own table has Opus narrowly ahead of Astra and Argon. The scores there are 59.6%, 59.1%, and 57.1%.
  • OSWorld-2.1, a test of using a computer: Opus leads that same table at 87.2%. Astra follows at 80.9%, and Argon at 74.7%.
  • Staying inside limits: Anthropic reports about 85% fewer attempts to break limits than Opus 5. That number comes from Anthropic alone. Running experiments in a closed, safe space is still wise.

Downsides exist. Opus cost the most per task at top effort. Like any model, Opus can invent sources and state wrong facts with confidence. That is why the human checks stay in the chain.

Gemini 4 Argon: documents and writing

Once results exist, the next need is a careful writer. That is Argon's job here. Three things make Argon a good fit:

  • Long output: Argon can write up to 1 million tokens in one job. That suits large documents.
  • Long input: Argon scored 84.2% on GraphWalks, a test of tracking links across long text. Astra scored 71.8%, and Opus scored 66.8%.
  • Legal work: Argon scored 19.6% on Harvey's legal test. Astra scored 5.4%, and Opus scored 3.8%.

One fair warning: Google's table has no test for writing quality. This choice rests mostly on personal use and the wrong-answer numbers below.

Argon has weak spots too. Argon trails Anthropic's newest models on long coding tasks. Bloomberg also reported internal doubts about real-world results matching the scores. A low rate of wrong answers does not make every source safe. Give Argon the sources, and ask Argon to use only those.

"Fewer wrong answers" is true, with a catch

Artificial Analysis runs a test called AA-Omniscience. The test asks factual questions and lets a model decline to answer. The table shows the results out of 100 questions. The correct column comes from reported accuracy. The other two columns are rough estimates from reported rates.

Out of 100 questions Correct Wrong Declined or partial
Gemini 4 Argon (High) 50 about 8 about 42
GPT-6 Astra (Max) 63 about 19 about 18
Claude Opus 5.5 66 about 20 about 14

So yes, Argon gives far fewer wrong answers than Astra. The reported rate of made-up answers was 15% for Argon and 51% for Astra. But Astra also gets more answers right. Much of Argon's lead comes from declining more questions.

That trade-off suits writing. In a report, an honest gap beats a confident mistake. During brainstorming, the same habit would be annoying.

Three limits apply. The Opus row comes from a separate reading at a different effort setting. So compare that row loosely. The test has also changed, and the test page now shows about 4% for Argon. So treat these figures as a snapshot in time. Finally, this test checks memory of facts. A summary of a private document needs a separate check.

Which model suits which research

Research area First pick Why Watch out for
Biology, bioinformatics Opus 5.5 Long analysis with human checks Scores disagree, and safety blocks can interrupt work
Defensive security Astra or Argon Stronger public scores, fewer default blocks Full access needs approval
Math and theory Astra Strong FrontierMath score, good at pushing back Argon wins on RiemannBench in Google's table
Machine learning experiments Opus 5.5 Leads the long-task tests above Cost per task
Law, finance, long documents Argon Leads the Harvey and Vals Finance tests Uses many tokens per task

Biology

For bioinformatics work, Opus is the first choice here. Calling Opus the king fits this way of working. Anthropic reports gains on a long molecular design test with Dyno Therapeutics. Anthropic also says Opus matches or beats Claude Mythos 5.1 in biology.

The public scores do not all agree, though:

  • LABBench2: One tracker lists Argon first at 88.8%. One comparison gives Astra 85.4% and Opus 73.1%. The test's own leaderboard lists none of these models yet. So treat these scores as unconfirmed.
  • Terminal-Bench-Science: Anthropic's own table has Astra ahead of Opus, 64.6% to 58.7%.
  • A new rival: OpenAI started rolling out GPT-6 Rosalind Research on 30 October. This life-science model goes only to approved organizations.

Both views can be true. LABBench2 tests separate skills, like reading papers, figures, and DNA sequences. Daily bioinformatics work is a long project where the plan keeps changing. Opus fits that style of work. That is personal experience, not a measured result. Opus also ships with strict biology safety blocks. Affected work can go through Anthropic's Life Sciences Verification Program.

Cybersecurity

The case here is stronger. This section covers only defensive work, like finding and fixing flaws.

  • Fixing flaws: On CWE-bench, the three nearly tie. Astra and Argon score 68%, and Opus scores 67%.
  • Harder tests: On ExploitGym, Astra and Argon score about 42%. Opus scores 34.5%. On SEC-bench Pro, Astra scores 85.4% and Argon 83.4%. No Opus score is reported.
  • Access: By default, Opus 5.5 sends most security tasks to Opus 4.8. Only verified users avoid this. For most people, that settles the question.

All three companies set limits here. Astra refused to write working attack code at launch. OpenAI plans wider access through the Daybreak program. Google's fullest versions go through the Fairwind Program. Anthropic also limits access to the strongest security model. So this comparison covers only what ordinary users get.

One more number matters for AI tools that read unknown web pages. Such pages can hide harmful instructions. On Gray Swan's test, these attacks worked 0.7% of the time on Argon. The rate was 1.0% on Opus and 8.5% on Astra.

Five rules for all three

The workflow matters more than the model. These rules stay the same for every model:

  1. Plan before running. No experiment starts until a human has read the plan.
  2. Use a second model as a critic. Argon reviews the findings from Opus. Astra hunts for holes in both.
  3. Open every source. A source nobody has read does not go in.
  4. Limit what code can touch. Give AI tools only the access the job needs.
  5. Record the model, version, and date. A result with no record is hard to defend.

What comes next: leaks and a forecast

The Final Road We SEE Now
Some of these paths are real products. Some may be only rumours.

Leaks are not facts. Names change, launches slip, and some models never ship. Everything in this section is a forecast, checked on 5 November. The timeline below shows the story so far.

timeline
    title New AI models, autumn 2026
    3 Sep : GPT-6 Astra
    22 Sep : Claude Opus 5.5
    28 Sep : GPT-6.1 Astra shelved
    30 Sep : Gemini 4 Argon announced
    28 Oct : Gemini 4 Carbon announced for security teams
    30 Oct : GPT-6 Rosalind Research rollout starts
    Next : Claude Fable 5.5 and GPT-6.2 Astra, if the leaks are right
Model Status on 5 Nov Main evidence Forecast Confidence
Gemini 4 Carbon Announced, limited access Google's own table Wider release soon, at Argon's price High
GPT-6.2 Astra Not announced Usage data, server reply, press report An upgrade for long coding and research Fairly high
Claude Fable 5.5 Not announced Price list entry, tester reports A new top model for long reasoning Medium
Claude Mythos 5.5 Not announced Hidden text on Anthropic's page A limited-access model for approved researchers Medium
Gemini 4 Xenon Unconfirmed One anonymous leak A cheaper, faster model, if Xenon exists Low

GPT-6.2 Astra: the freshest leak

OpenAI has announced nothing, but the signs keep growing:

  • Usage reports: Users of Codex, OpenAI's coding tool, spotted the name gpt-6.2-astra. The model menu still says GPT-6 Astra.
  • Share of work: Reports range from 1.6% of a week to 37.7% of a day.
  • Server reply: One user shared a raw server reply carrying the same name.
  • A second name: An unverified screenshot also shows gpt-6.2-betelgeuse.
  • A possible date: A post citing The Information says launch could come by 6 November. That post also mentions Astra, Sol, and Cyber versions. The original report sits behind a paywall, so the date stays unconfirmed.

Sam Altman added fuel on 2 November. Speaking in Abu Dhabi, Altman claimed a Millennium Prize math result. Altman said 10,000 AI helpers ran an upcoming GPT model for about four days. The newspaper could not verify that claim. People at the event also said the next GPT would come soon.

History matters here too. OpenAI shelved GPT-6.1 Astra in late September over safety concerns. So 6.2 may be a retrained version, though that link is a guess. The likely result is an upgrade, not a new generation. Factual accuracy is the number to watch. Better accuracy would make Astra useful later in the chain.

Claude Fable 5.5 and Mythos 5.5: strong signals, no facts

Anthropic has announced nothing here either. These are the main leaks:

Two other claims look weak. "Claude Next" rests on one social media post. "Sonnet 5.5 Extra" looks like Sonnet 5.5 at a higher effort setting.

Fable 5.1 and Mythos 5.1 launched together, so a paired launch seems likely. Anthropic is also preparing a stock market listing. Investor meetings can start from 10 November. That may affect timing, though no source proves a link.

Opus 5.5 already beats Fable 5.1 on most public tests. That leaves Fable 5.1 hard to justify at $10 / $50. A new Fable needs a clear lead, most likely on long reasoning tasks. Biology would be the first thing to test.

Gemini 4 Carbon: already real

Carbon is the only real product in this list. Google says Carbon is best on 15 of 17 tests in the company's own table. Carbon scores 59 on the Artificial Analysis index. Opus scores 58, Astra 57, and Argon 50. Google also says Carbon cuts unsupported claims by about half compared with Argon.

These numbers come from Google, and only approved security teams have access so far. Testers also report a newer build on Arena, a public model testing site. If independent tests agree, Carbon will likely take over the writing job. Carbon may challenge Opus for experiments too.

Gemini 4 Xenon: the weakest rumour

Xenon deserves the most doubt. The only source found is one article citing unnamed insiders. That article claims a faster model at up to 50% lower token prices. The article shows no test scores. Several more searches found nothing else. Google's model page names only Argon and Carbon. Plans should not depend on Xenon yet.

The short version

Do not hunt for one winner. Match the model to the job, and keep a person at every decision point.

Today, that means Astra for ideas, Opus for plans and experiments, and Argon for writing. At least one of those will likely change within months, probably with Carbon. The chain should outlast every model.