Back to the trend of the day: Google releases its smartest AI yet, and 10,000 scientists get it free
Poly Rating · from the trend of the day

Best AI for Scientific Research: Top 10 Models, 2026

Anthropic gives academic PIs a free Claude seat for a year and Google gives 10,000 selected scientists six months of Argon's top tier, while OpenAI's offer requires a paper posted in the last three years.

  • No. 1: Gemini 4 Argon
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    Gemini 4 Argon

    Google gives 10,000 selected scientists its top Antigravity tier free for six months, half Anthropic's term. Index 53, and second on Humanity's Last Exam at 57.1%.

    Included it: Claude, ChatGPT.

    In their lists: Claude: 2 · ChatGPT: 1

    What each AI noted 2
    • Claudeplace 2

      Google's offer is 10,000 scientists, top Antigravity tier, six months, free. That is half the term of Anthropic's, with the highest usage tier and no card.

      Source ↗
    • ChatGPTplace 1

      Google offers selected scientists six months of top-tier access at no charge, while Artificial Analysis reports 53 and $1.99 per index run.

      Source ↗
    2 AIs put it here
    Source ↗
  2. 2

    Claude Opus 5.5

    Highest independent index (58) plus the longest free term: a free standard Claude Team seat for a year for PIs at academic or nonprofit institutions.

    Included it: Claude, ChatGPT.

    In their lists: Claude: 1 · ChatGPT: 3

    What each AI noted 2
    • Claudeplace 1

      The highest independent index (58) plus the longest free-access term of the three big programs: a free standard Claude Team seat for a year, with 10,000 seats and 5x-limit premium seats at $15 a month.

      Source ↗
    • ChatGPTplace 3

      Artificial Analysis identifies it as the index leader, but index leadership is not evidence that it produces publishable science.

      Source ↗
    2 AIs put it here
    Source ↗
  3. 3

    GPT-6 Astra

    Top GPQA Diamond score at 96.3%, but OpenAI's program requires a paper posted in the last three years, and its only report names GPT-5.6 models, not Astra.

    Included it: Claude, ChatGPT.

    In their lists: Claude: 3 · ChatGPT: 2

    What each AI noted 2
    • Claudeplace 3

      The strongest raw science score found: 96.3% on GPQA Diamond. Access is the narrowest of the three: you must have posted a paper in the past 3 years.

      Source ↗
    • ChatGPTplace 2

      Artificial Analysis reports 53 on its broad index; the sources retrieved do not verify an equivalent six-month science-access offer.

      Source ↗
    2 AIs put it here
    Source ↗
  4. More about the top three ↓
  5. 4

    Claude Fable 5.1

    Leads Humanity's Last Exam on AA's table at 59.1%, but costs $10/$50 per million tokens, and no scientist program was found that names it.

    Included it: Claude, ChatGPT.

    In their lists: Claude: 4 · ChatGPT: 4

    What each AI noted 2
    • Claudeplace 4

      The top Humanity's Last Exam score on AA's table (59.1%), but it has the highest list price of the leaders at $10/$50, and I found no scientist-program terms naming it.

      Source ↗
    • ChatGPTplace 4

      An October roundup places it among leading models on several boards, but the material retrieved does not confirm a science-specific result or open research program.

      Source ↗
    2 AIs put it here
    Source ↗
  6. 5

    Gemini 3.8 Flash

    Scores 95.3% on GPQA Diamond, within a point of Astra, from a smaller model with a broad index of 41. No researcher-access terms were found.

    Included it: Claude, ChatGPT.

    In their lists: Claude: 8 · ChatGPT: 8

    What each AI noted 2
    • Claudeplace 8

      At 95.3% on GPQA Diamond (AA, high effort) it is within a point of the leader. It is a smaller model, and I found no pricing or access terms for it.

      Source ↗
    • ChatGPTplace 8

      Its broad index score is 33; ranking it for science is necessarily tentative without a verified science-specific result.

      Source ↗
    2 AIs put it here
    Source ↗
  7. 6

    Claude Sonnet 5.5

    Index 56 at $2/$10, the practical pick for bulk literature and data work. Anthropic's scientist seats apply, though no model-by-model list was found.

    Included it: Claude. Did not include: ChatGPT.

    In their lists: Claude: 5

    What each AI noted 1
    • Claudeplace 5

      Index 56 at $2/$10, so it is the model to use for bulk literature and data work. It follows the same Anthropic seat and API-credit terms as Opus 5.5 where the program covers it.

      Source ↗
    1 AIs put it here
    Source ↗
  8. 7

    Grok 4.7

    Broad index 46. No science-specific score and no research-access terms were verified.

    Included it: ChatGPT. Did not include: Claude.

    In their lists: ChatGPT: 5

    What each AI noted 1
    • ChatGPTplace 5

      Artificial Analysis lists a broad index score of 46; no verified research-access term was found in the sources gathered.

      Source ↗
    1 AIs put it here
    Source ↗
  9. 8

    GPT-6.1 Sol

    The cheapest high-end reasoner to run, at $0.72 a task with index 52, but no science-specific score and no matching scientist program were found.

    Included it: Claude. Did not include: ChatGPT.

    In their lists: Claude: 6

    What each AI noted 1
    • Claudeplace 6

      Index 52 at $0.72 a task is the cheapest high-end science reasoner to run, but I found no science-specific score for it.

      Source ↗
    1 AIs put it here
    Source ↗
  10. 9

    Muse Wonder (Meta)

    A closed-weight preview at index 52 and $4/$20 with no free researcher program found, so it ranks on tested ability alone.

    Included it: Claude. Did not include: ChatGPT.

    In their lists: Claude: 7

    What each AI noted 1
    • Claudeplace 7

      A new closed-weight preview at index 52 and $4/$20 with no research-access scheme found, so it ranks on tested ability only.

      Source ↗
    1 AIs put it here
    Source ↗
  11. 10

    Claude Haiku 5.5

    The lowest-cost option for large-volume screening, at $0.10/$0.50 per million tokens (prompts up to 100K). No independent science score was found.

    Included it: Claude. Did not include: ChatGPT.

    In their lists: Claude: 10

    What each AI noted 1
    • Claudeplace 10

      Released Oct 7 at $0.10/$0.50 per million tokens (up to 100K-token prompts). It is the lowest-cost option for large-volume screening, with no independent science score found.

      Source ↗
    1 AIs put it here
    Source ↗

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

Index scores are comparable, but the free offers are not: six months for Google, a year for Anthropic, and only recent paper authors for OpenAI.

1Gemini 4 Argon

No. 1 — Gemini 4 Argon. On the lists of 2 of 2 AIs; median place there: 1.5. AIs liked its free six-month top tier for 10,000 scientists and its strong second place on Humanity's Last Exam.

Did you know

  • Gemini 4 Argon can write a very long answer. One report says it takes text, images and video in, and can send text back, with up to 1 million tokens on each side. That is roughly a whole shelf of books.

    Source ↗
What it is
Google's top AI model, offered to scientists through its Antigravity tool.
Its edge
Second on Humanity's Last Exam, and a free offer open to selected scientists.
Who it suits
Suits scientists Google selects and tasks with long texts; not those needing top coding.
  • Index on Artificial AnalysisIndex 53
Why it is here

  • On Artificial Analysis’s independent index, Gemini 4 Argon scores 53, the same as GPT-6 Astra and 5 points below Claude Opus 5.5. — (voice only)

    Source ↗

  • On Humanity’s Last Exam, Gemini scores 57%; that is 2 points above GPT-6 Astra’s 55%, but 4 below Claude Opus 5.5’s 61%. — (voice only)

    Source ↗
Did you know

  • Argon is described as a low-hallucination model built for knowledge work. That means it is tuned to make up fewer false facts, which matters when a researcher is checking sources.

    Source ↗
History

  • In 2024, Gemini made historically inaccurate images of people, and Google paused that feature while it worked on a fix. It was a public setback for the model family. — (voice only)

    Source ↗
Keep in mind

  • Tests disagree. Argon ranks first on Vals AI's professional-work index but eighth of 216 on the Artificial Analysis Intelligence Index, behind Claude Opus 5.5. Coding is a weak spot: 57.4% on Terminal-Bench 4.0 against Opus 5.5's 66.4%.

    Source ↗

  • The $2 per million input tokens price lasts only through November 30. On December 1 both rates double. Antigravity needs an AI Ultra subscription for individuals, and Google's own productivity figures haven't been independently replicated.

    Source ↗
What people say

  • "This is a really, really, really good model. But is it the best model out there? I actually don't think it is." — Sam Witteveen, Red Dragon AI (to VentureBeat)

    Source ↗
Where to go further

  • Look first at the offer details: six months of free access to Antigravity's highest usage tier, granted on a rolling basis after selection. The article gives no sign-up link, so check Google's Antigravity site before applying.

    Source ↗
Compare the top three ↓

2Claude Opus 5.5

No. 2 — Claude Opus 5.5. On the lists of 2 of 2 AIs; median place there: 2. It has the highest independent index and the longest free term: a full year of a Claude Team seat for academic PIs.

Did you know

  • Claude Opus 5.5 costs less than the model before it. Per million tokens, input fell from $5 to $4 and output from $25 to $20. Anthropic says it runs about 40% cheaper than Opus 5 at default settings.

    Source ↗
What it is
Anthropic's newest top Claude model, built for hard reasoning and coding.
Its edge
Highest index and best Humanity's Last Exam score, plus the longest free access.
Who it suits
Suits academic or nonprofit lab leaders wanting a year free; not people outside those institutions.
  • Index on Artificial AnalysisIndex 58
Why it is here

  • Claude Opus 5.5 scores 58 on Artificial Analysis’s independent index, 5 points ahead of Gemini 4 Argon and GPT-6 Astra, both at 53. — (voice only)

    Source ↗

  • It scores 61% on Humanity’s Last Exam, ahead of Gemini 4 Argon at 57% and GPT-6 Astra at 55%. — (voice only)

    Source ↗
Did you know

  • Opus 5.5 cannot switch its reasoning off. It also hands most cybersecurity tasks to an older model, Opus 4.8, while ordinary software bug fixing stays allowed.

    Source ↗
History

  • In 2026, Anthropic called for “pacing the frontier”; Opus 5.5 was its first release after that call. The launch paired its most capable Opus line with safeguards similar to those on its top model. — (voice only)

    Source ↗
Keep in mind

  • Some Reddit users say Opus 5.5 got worse about a week after launch. The posts are anecdotes, and nothing in the threads confirms Anthropic changed anything. Commenters also question how the sentiment tracker measures this.

    Source ↗

  • Opus 5.5 is not the cheap choice. After Argon's introductory price ends on December 1, both models sit in the same per-token price tier.

    Source ↗
Where to go further

  • Open this page first to see who qualifies: principal investigators at accredited universities and nonprofit research institutes. Standard seats are free for 12 months; Premium seats are $15 per user per month.

    Source ↗
Claude put it first. What they saw in it

  • Claude: The highest independent index (58) plus the longest free-access term of the three big programs: a free standard Claude Team seat for a year, with 10,000 seats and 5x-limit premium seats at $15 a month.

Compare the top three ↓

3GPT-6 Astra

No. 3 — GPT-6 Astra. On the lists of 2 of 2 AIs; median place there: 2.5. It has the top GPQA Diamond score, but the narrowest access: OpenAI requires a paper posted in the last three years.

Did you know

  • GPT-6 Astra is OpenAI's first launch with critical cybersecurity capability. It first went to customers in the Daybreak program. OpenAI said it was trained on more than 100,000 GPUs at its Stargate site in Texas.

    Source ↗
What it is
OpenAI's newest flagship model, rolled out in stages.
Its edge
Highest raw science-question score (96.3% on GPQA Diamond) of the three.
Who it suits
Suits researchers with a recent paper and budget; not free-tier users or audio and video work.
  • Index on Artificial AnalysisIndex 53

Wikipedia · Wikidata

Why it is here

  • On Artificial Analysis’s independent index, Astra scores 53, level with Gemini 4 Argon and 5 points below Claude Opus 5.5. — (voice only)

    Source ↗
History

  • In 2026, a Hugging Face incident shaped Astra’s safety testing: OpenAI said it built a new test to check whether the model would exceed its intended scope on difficult or impossible tasks. — (voice only)

    Source ↗
Keep in mind

  • Astra's list price is double GPT-5.6 Sol's on input and output. Requests over 272K input tokens are billed at higher rates for the whole request, and fast mode doubles the rates again.

    Source ↗

  • Rollout is staged, starting with trusted organizations, and the free tier is not included. Audio and video are unsupported. The article notes OpenAI's benchmark gains are vendor claims that have not been independently verified.

    Source ↗
Where to go further

  • Open OpenAI's GPT-6 Astra page and read the computer use and coding sections first. Then open the safety overview page linked from it to see how OpenAI describes its cybersecurity limits.

    Source ↗
Compare the top three ↓

The top three side by side

Gemini 4 ArgonClaude Opus 5.5GPT-6 Astra
In the AIs' lists2 / 22 / 22 / 2
Index on Artificial AnalysisIndex 53Index 58Index 53
Put first byChatGPTClaude—

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

More ratings

03Worth knowing

What the AIs noticed while building the list.

Three makers each offer about 10,000 free scientist places in 2026 on different terms: Anthropic one year for PIs (Aug 27), Google six months by selection (Oct 13), and OpenAI a first cohort of 10,000 out of 100,000 that requires a recent paper (July 29).

Noticed by:
Included it: Claude, ChatGPT.
Source ↗

The science leader depends on the test. Astra tops GPQA (96.3%). On Humanity's Last Exam, Fable 5.1 leads AA's table (59.1%), Opus 5.5 leads Klu's text-only board (61.4%) and Astra leads Scale's (54.2%).

Noticed by:
Included it: Claude. Did not include: ChatGPT.
Source ↗

Argon's low 15% hallucination rate comes partly from answering less: its accuracy on hard factual questions is 50% against Astra's 63%. Keep that in mind for literature work.

Noticed by:
Included it: Claude. Did not include: ChatGPT.
Source ↗

GPQA and FrontierMath test expert science and research-level math questions. Neither shows that a model can carry out a research project or produce publishable findings.

Noticed by:
Included it: ChatGPT. Did not include: Claude.
Source ↗

Anthropic's API-credit amount differs by source: $20,000 on the program page, $30,000 for a 2026 cycle that closed July 15, and $50,000 on one third-party blog.

Noticed by:
Included it: Claude. Did not include: ChatGPT.
Source ↗

04Questions people ask

what is the best ai for scientific research

On the independent index, Claude Opus 5.5 leads at 58, ahead of Sonnet 5.5 (56) and Argon (53), and it comes with a free one-year seat for academic PIs. On GPQA, GPT-6 Astra leads at 96.3%.

Source ↗
what is the best ai for science research

Tested science results are split: Astra is first on GPQA at 96.3%, Fable 5.1 is first on AA's Humanity's Last Exam table at 59.1%, and Argon is second there at 57.1%. Sources differ because the boards use different scoring modes.

Source ↗
what is the best ai for scientific research and writing

No independent science-writing score was found. Argon is first in the Arena human vote (about 1524), and Opus 5.5 leads the independent index (58).

Source ↗
what is the best free ai for scientific research

Anthropic gives academic PIs a free standard Claude seat for a year. Google offers 10,000 selected scientists six months of Argon's top Antigravity tier. All such programs are selective, not open to everyone.

Source ↗

05How we counted

In plain words, no formulas.

1The question

Top 10 AI Models Researchers Can Actually Use for Science

2Why this topic

The reader gets tested science ability next to the practical question of who can use the model, for how long and on what terms, which no single benchmark table shows.

3Where it applies

Global scientific research and research support

4What we looked at

  • Independent performance on science and reasoning tasks, alongside broad capability scores
  • Access terms: who is eligible, for how long, and whether research use is free or paid
  • Artificial Analysis Intelligence Index and per-evaluation results — Independent multi-task index with a scientific-reasoning category (CritPt, 70 tasks, 10% of the index). It does not test whether a m
  • Humanity's Last Exam — Expert-level academic questions. The leader depends on the board and scoring mode (text-only, with tools).
  • GPQA Diamond — 198 PhD-level science multiple-choice questions written by experts. It tests science question answering, not end-to-end research.
  • Google DeepMind launch results — Science tests Google reported itself: LabBench and RareBench (private evals), and Terminal-Bench Science in Google's rival table.
  • Scientist access programs (Anthropic, Google, OpenAI) — Who qualifies, for how long and at what price, from each maker's announcement. These are program terms, not benchmark scores.

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT.

6How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

06Each AI's own list

🇺🇸 Claude10 places · 3 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    Claude Opus 5.5

    The highest independent index (58) plus the longest free-access term of the three big programs: a free standard Claude Team seat for a year, with 10,000 seats and 5x-limit premium seats at $15 a month.

    Source ↗
  2. 2
    Gemini 4 Argon (Google, 2026)

    Google's offer is 10,000 scientists, top Antigravity tier, six months, free. That is half the term of Anthropic's, with the highest usage tier and no card.

    Source ↗
  3. 3
    GPT-6 Astra (OpenAI)

    The strongest raw science score found: 96.3% on GPQA Diamond. Access is the narrowest of the three: you must have posted a paper in the past 3 years.

    Source ↗
  4. 4
    Claude Fable 5.1

    The top Humanity's Last Exam score on AA's table (59.1%), but it has the highest list price of the leaders at $10/$50, and I found no scientist-program terms naming it.

    Source ↗
  5. 5
    Claude Sonnet 5.5

    Index 56 at $2/$10, so it is the model to use for bulk literature and data work. It follows the same Anthropic seat and API-credit terms as Opus 5.5 where the program covers it.

    Source ↗
  6. 6
    GPT-6.1 Sol

    Index 52 at $0.72 a task is the cheapest high-end science reasoner to run, but I found no science-specific score for it.

    Source ↗
  7. 7
    Muse Wonder (Meta)

    A new closed-weight preview at index 52 and $4/$20 with no research-access scheme found, so it ranks on tested ability only.

    Source ↗
  8. 8
    Gemini 3.8 Flash (Google)

    At 95.3% on GPQA Diamond (AA, high effort) it is within a point of the leader. It is a smaller model, and I found no pricing or access terms for it.

    Source ↗
  9. 9
    Claude Opus 5

    Still available at $5/$25, and index 50.8 puts it on the science shortlist only where a lab is already on it.

    Source ↗
  10. 10
    Claude Haiku 5.5

    Released Oct 7 at $0.10/$0.50 per million tokens (up to 100K-token prompts). It is the lowest-cost option for large-volume screening, with no independent science score found.

    Source ↗

What Claude noticed

Three makers each announced about 10,000 free seats for scientists in 2026, with different terms. Anthropic's is one year (Aug 27), Google's is six months (Oct 13), and OpenAI's first cohort is 10,000 of 100,000 with a paper requirement (July 29).

Source ↗

Science leaders vary by test. GPT-6 Astra tops GPQA (96.3%), Fable 5.1 tops HLE on AA's table (59.1%), Opus 5.5 tops HLE text-only on Klu (61.4%), and Argon is second on HLE (57.1%) and has the lowest hallucination rate (15%).

Source ↗

Argon's 15% hallucination rate comes partly from answering less. Its accuracy on hard factual questions is 50% against Astra's 63%, a caution for literature work.

Source ↗
🇺🇸 ChatGPT10 places · 4 notes

The same question, answered by ChatGPT alone. The final list above combines all the AIs.

  1. 1
    Gemini 4 Argon

    Google offers selected scientists six months of top-tier access at no charge, while Artificial Analysis reports 53 and $1.99 per index run.

    Source ↗
  2. 2
    GPT-6 Astra

    Artificial Analysis reports 53 on its broad index; the sources retrieved do not verify an equivalent six-month science-access offer.

    Source ↗
  3. 3
    Claude Opus 5.5

    Artificial Analysis identifies it as the index leader, but index leadership is not evidence that it produces publishable science.

    Source ↗
  4. 4
    Claude Fable 5.1

    An October roundup places it among leading models on several boards, but the material retrieved does not confirm a science-specific result or open research program.

    Source ↗
  5. 5
    Grok 4.7

    Artificial Analysis lists a broad index score of 46; no verified research-access term was found in the sources gathered.

    Source ↗
  6. 6
    Gemini 3.8 Flash (High)

    Artificial Analysis reports 41; that broad capability score cannot stand in for a graduate science benchmark.

    Source ↗
  7. 7
    Gemini 3.8 Flash (Medium)

    The index score is 40; the retrieved material does not establish science-task scores or a dedicated research-access offer.

    Source ↗
  8. 8
    Gemini 3.8 Flash (Low)

    Its broad index score is 33; ranking it for science is necessarily tentative without a verified science-specific result.

    Source ↗
  9. 9
    Gemini 3.1 Pro Preview

    Artificial Analysis reports 30, far below Argon’s 53 on the same listed index; neither number measures real-world discovery.

    Source ↗
  10. 10
    Other models with public science scores

    The retrieved sources do not support a responsible tenth model placement with verified access duration and science-task results; use the listed science leaderboards to build a complete shortlist.

What ChatGPT noticed

GPQA Diamond has 198 questions in its highest-quality subset, but a strong result still measures science question answering rather than whether a system can carry out a full research project.

Source ↗

FrontierMath uses research-level problems and automated verification, but success on it does not establish that a model generates publishable scientific findings.

Source ↗

Google’s scientist offer is time-limited and selective: up to 10,000 applicants can receive six months of the highest Antigravity usage tier, subject to eligibility, capacity and selection.

Source ↗

Artificial Analysis’s broad index includes scientific reasoning among multiple evaluation areas; the index is useful for context but does not isolate science or research writing.

Source ↗