Home › AI Benchmark › Claude Opus 4.7
Anthropic released on Apr 16, 2026 API access

Claude Opus 4.7

Claude Opus 4.7 shipped on Apr 16, 2026 as the most capable model Anthropic had made generally available, and 42 days later Opus 4.8 had taken its place at the same price. It is now legacy: it costs US$ 5 per million input tokens and US$ 25 per million output, against US$ 4 and US$ 20 for Opus 5.5, which the documentation says to migrate to, and it will not be retired before Apr 16, 2027. It is the model that introduced the tokenizer, the xhigh effort level and the high-resolution images that later models inherited.

#26 of 267 in the ECI ranking
156.32 ECI, the Epoch AI capability index
1,532 Arena ELO, human preference · #16
76.5 LiveBench, overall average from 0 to 100 · #23
No votes yet. Rate it.
Spec sheet

Claude Opus 4.7 in official numbers

Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.

Maker
Anthropic
API identifier
claude-opus-4-7
Release
Apr 16, 2026
Status
legacy; the documentation says to migrate to Opus 5.5
Input
US$ 5 per million tokens
Output
US$ 25 per million tokens
Cache read
US$ 0.50 per million (10% of input; US$ 0.20 on Opus 5.5)
Cache write
US$ 6.25 (5 minutes) or US$ 10 (1 hour)
Batch API
US$ 2.50 input and US$ 12.50 output
Fast mode
not available; the request returns an error
Context window
1 million tokens, no surcharge
Max output
128 thousand tokens (300 thousand in batch, beta)
Reliable knowledge
through January 2026
Reasoning
adaptive, off unless the request turns it on; default effort high
Cache minimum
2,048 tokens (512 on Opus 5.5)
Images
up to 2,576 pixels on the long edge
Where it runs
Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS
Capacity
supports Priority Tier
Retirement
not before Apr 16, 2027
Analysis

Where Claude Opus 4.7 does well, where it falls behind and who it is for

Where it does well

  • Coding: 64.3% on SWE-bench Pro at launch, against 57.7% for GPT-5.4, 54.2% for Gemini 3.1 Pro and 53.4% for Opus 4.6.
  • Computer use: 78.0% on OSWorld-Verified, against 75.0% for GPT-5.4, with images up to 2,576 pixels on the long edge, more than three times the pixels of earlier models.
  • Tools at scale: 77.3% on MCP-Atlas, ahead of Gemini 3.1 Pro (73.9%) and GPT-5.4 (68.1%).
  • Office and finance work: 1,753 on GDPval-AA, against 1,674 for GPT-5.4, and 64.4% on Finance Agent v1.1, against 61.5% for GPT-5.4 Pro.
  • Prompt injection: according to the Opus 4.8 system card, its successor was somewhat less robust than 4.7 in agentic contexts.

Where it falls behind

  • Web search: 79.3% on BrowseComp, below Opus 4.6 itself (83.7%) and GPT-5.4 Pro (89.3%).
  • Terminal: 69.4% on Terminal-Bench 2.0, against 75.1% for GPT-5.4, which OpenAI measured with its own harness.
  • No fast mode: a request with speed: "fast" returns an error; Opus 4.8, Opus 5 and Opus 5.5 have the mode.
  • Demanding cache: it only caches prompts from 2,048 tokens, against 512 on Opus 5.5, and rereads for US$ 0.50, against US$ 0.20.
  • More tokens for the same text: 1.0 to 1.35 times as many tokens as on Opus 4.6, at the same price per token.

Who it makes sense for

  • Systems validated on it that need time to rerun tests: it will not be retired before Apr 16, 2027, with at least 60 days' notice.
  • Accounts with a Priority Tier contract, which Opus 4.7 supports and Opus 5.5 does not.
  • Pipelines that run without thinking, force a tool with tool_choice or use computer_20251124: all of that works on 4.7 and returns an error on Opus 5.5.
  • Anyone who needs to reproduce a result measured on it, such as an old internal eval.

When to pick another model

  • New projects: the documentation says to start with Opus 5.5, at US$ 4 and US$ 20 per million tokens.
  • Staying on 4.7 only for compatibility: Opus 4.8 supports thinking off, forced tool use and Priority Tier at the same price, has fast mode and will not be retired before May 28, 2027.
  • Long instructions repeated across many requests: the 2,048-token cache minimum and the US$ 0.50 reread add up.
  • Tasks that need recent knowledge: its reliable cutoff is January 2026; Opus 5.5's is June 2026.

What Anthropic published at launch

Opus 4.7 arrived on Apr 16, 2026 as a direct upgrade to Opus 4.6, at the same price. Anthropic highlighted hard coding work, high-resolution vision and instructions followed to the letter, and placed it below Claude Mythos Preview, which had restricted access.

The table below is the one from the announcement, without the Mythos Preview column. In the ten rows we picked, Opus 4.7 leads five, GPT-5.4 leads four and Opus 4.6 takes one. Caveats from Anthropic itself: the OpenAI column uses the best version available on the API, sometimes GPT-5.4 Pro, and on Terminal-Bench OpenAI used its own harness.

Opus 4.7 in the figures Anthropic published at launch
Opus 4.7
Anthropic		Opus 4.6
Anthropic		GPT-5.4
OpenAI		Gemini 3.1 Pro
Google
SWE-bench Pro
coding		 64.3%53.4%57.7%54.2%
Terminal-Bench 2.0
terminal coding		 69.4%65.4%75.1%68.5%
Humanity's Last Exam
without tools		 46.9%40.0%42.7%44.4%
Humanity's Last Exam
with tools		 54.7%53.3%58.7%51.4%
BrowseComp
web search		 79.3%83.7%89.3%85.9%
MCP-Atlas
tools at scale		 77.3%75.8%68.1%73.9%
OSWorld-Verified
computer use		 78.0%72.7%75.0%not published
Finance Agent v1.1
financial analysis		 64.4%60.1%61.5%59.7%
GPQA Diamond
graduate-level science		 94.2%91.3%94.4%94.3%
CyberGym
reproducing security flaws		 73.1%73.8%66.3%not published

Figures from the Claude Opus 4.7 announcement, checked against table 8.1.A of the system card (max effort, averaged over 5 trials). The dot marks the best in each row. In the GPT-5.4 column, the HLE, BrowseComp, Finance Agent and GPQA scores are GPT-5.4 Pro. At the Opus 4.8 launch, Anthropic changed how it runs OSWorld-Verified and remeasured Opus 4.7: the table shows 82.8%, and the footnote says 82.3%.

What Opus 4.7 changed and later models inherited

Much of what is now standard on the Anthropic API started with Opus 4.7, which shortens the jump for anyone still running it:

  • New tokenizer: the same text becomes 1.0 to 1.35 times as many tokens as on Opus 4.6. Opus 4.8, Opus 5 and Opus 5.5 use the same tokenizer.
  • xhigh effort: between high and max, for agentic and coding tasks of more than 30 minutes.
  • High-resolution images: up to 2,576 pixels on the long edge, against 1,568 before. Each full-resolution image can use up to about 3 times as many image tokens.
  • Removed parameters: non-default temperature, top_p and top_k and the manual thinking budget started returning a 400 error.
  • Thinking hidden by default: thinking blocks arrive empty unless the request asks for the summary.
  • Cybersecurity safeguards: it brought to the Opus line the safeguards that block high-risk requests in real time.

Is Claude Opus 4.7 still worth using in 2026?

As your main model, no. Opus 4.8 shipped 42 days later at the same price and, in its own launch table, beat 4.7 on every row: 69.2% against 64.3% on SWE-Bench Pro and 1,890 against 1,753 on GDPval-AA. Then came Opus 5 and Opus 5.5, which now costs 20% less per token and rereads the cache for US$ 0.20, against US$ 0.50. Anthropic has not published a table putting 4.7 side by side with Opus 5.5.

The reasons to hold on to 4.7 are a Priority Tier contract and code that runs without thinking, forces a tool or uses the old computer use tool. For those cases, Opus 4.8 supports the same at the same price, with fast mode and caching from 1,024 tokens. 4.7 stays available until at least Apr 16, 2027 on Anthropic's platforms; Amazon Bedrock and Google Cloud set their own dates.

Opus 4.7 against its successors, price per million tokens
Opus 4.7
legacy		Opus 4.8
legacy		Opus 5.5
current
Input US$ 5US$ 5US$ 4
Output US$ 25US$ 25US$ 20
Cache read US$ 0.50US$ 0.50US$ 0.20
Fast mode not availableUS$ 10 and US$ 50US$ 8 and US$ 40
Reasoning
when not requested		 offoffalways on
Default effort highhighmedium
Cache minimum 2,048 tokens1,024 tokens512 tokens
Priority Tier supportedsupportednot supported
Reliable knowledge January 2026January 2026June 2026
Retirement
not before		 Apr 16, 2027May 28, 2027Sep 22, 2027

Anthropic documentation: pricing, each model's page, prompt caching, migration guide and model deprecations, checked on Oct 1, 2026. Opus 5, also legacy, has the same prices as Opus 4.8.

What changes when you migrate to Opus 5.5

The migration guide treats Opus 4.7 like Opus 4.8. On Opus 5.5, this changes:

  • Thinking always on: a request without a thinking field now thinks, and turning it off returns a 400 error. Thinking tokens are billed as output.
  • Effort recalibrated: the default drops from high to medium. Rerun your effort sweep; at xhigh or max, start max_tokens at 64 thousand.
  • Forced tool use (any or tool) and computer_20251124 on the API and Google Cloud now return a 400 error.
  • Text between tool calls now arrives in thinking blocks, empty by default.
  • You gain fast mode, mid-conversation system messages (which 4.7 rejects with an error) and caching from 512 tokens; you lose Priority Tier.
In practice

Examples of using Claude Opus 4.7

An agent that reads dense screens

A computer use agent navigates systems with crowded screens. With images up to 2,576 pixels, Opus 4.7 scored 79.5% on ScreenSpot-Pro without tools, against 57.7% for Opus 4.6. Cost goes up too: each full-resolution image can use up to about 3 times as many image tokens.

An old prompt followed to the letter

Instructions Opus 4.6 followed loosely or skipped are followed literally on 4.7, and the output changes. Anthropic recommended retuning prompts and harnesses before switching.

A test before migrating

Run ten real tasks on Opus 4.7 at xhigh, the starting point Anthropic recommends for coding, and on Opus 5.5 at medium, its default. Compare accuracy, output tokens and cost per task.

What the tests say

Where Claude Opus 4.7 stands out and where it falls behind, by the numbers

Claude Opus 4.7 has published results on 37 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Jul 20, 2025 and Jul 15, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.

Where it stands out

Tests where it sits in the top third of its own generation.

Vending Bench 210,937
#1 of 43 in its generation · #5 of 57 overall
SWE-bench Verified83.5%
#1 of 26 in its generation · #1 of 32 overall
WebDev Arena1,567
#4 of 55 in its generation · #27 of 82 overall
Terminal-Bench80.2%
#3 of 31 in its generation, tied with 1 · #3 of 33 overall
AIME 2024 and 2025 (OTIS)97.8%
#8 of 69 in its generation · #28 of 168 overall

Where it falls behind

Tests where it sits in the bottom half of its own generation.

ForecastBench59.90
#19 of 30 in its generation · record: 62.50 (o3)
Furniture Assembly33.3%
#9 of 14 in its generation · record: 83.3% (Claude Opus 5.5)
Blueprint Bench 224.5%
#10 of 18 in its generation · record: 54.4% (Gemini 4 Argon)
Posttrainbench28.6%
#5 of 9 in its generation · record: 41.8% (Claude Fable 5)
See all 37 tests, one by one
TestScoreIn its generationOverallTest record
GPQA Diamond 90.2% #20 / 69 #38 / 173 96.4% Gemini 4 Argon
AIME 2024 and 2025 (OTIS) 97.8% #8 / 69 #28 / 168 100.0% GPT-5.5
Dtbench 94.7% #9 / 66 #21 / 148 98.9% Claude Opus 5.5
Lmca 52.20 #9 / 64 #18 / 121 68.23 Claude Opus 5.5
CritPt 12.0% #19 / 57 #45 / 112 32.3% GPT-5.6 Sol
WeirdML 76.4% #9 / 58 #14 / 112 93.6% GPT-6 Astra
SciCode 54.5% #9 / 53 #30 / 104 66.9% Claude Opus 5.5
Frontiermath Tiers 1 3 V2 70.2% #11 / 47 #26 / 86 93.7% GPT-6 Astra
ARC-AGI 2 75.8% #8 / 40 #16 / 83 95.0% GPT-6 Astra
ALE-Bench 1,323 #11 / 55 #20 / 82 2,951 GPT-6 Astra
WebDev Arena 1,567 #4 / 55 #27 / 82 1,820 Claude Opus 5.5
ARC-AGI 93.5% #10 / 40 #20 / 81 98.5% Claude Fable 5
SimpleQA Verified 51.7% #11 / 47 #25 / 81 76.0% Gemini 4 Argon
Chess puzzles 41.0% #18 / 44 #32 / 80 76.0% GPT-6 Astra
Game puzzles 28.0% #11 / 43 #32 / 77 84.0% GPT-6 Astra
ProofBench 54.0% #8 / 42 #26 / 74 100.0% Claude Fable 5.1
SimpleBench 61.7% #15 / 41 #19 / 74 81.9% Claude Fable 5
Frontiermath Tier 4 V2 31.7% #12 / 37 #27 / 67 100.0% GPT-6.1 Sol
FrontierMath 43.8% #6 / 38 #6 / 65 51.7% GPT-5.5
ForecastBench 59.90 #19 / 30 #30 / 62 62.50 o3
Vending Bench 2 10,937 #1 / 43 #5 / 57 15,515 GPT-6 Astra
Apex Agents 49.2% #8 / 24 #27 / 50 82.4% Gemini 4 Argon
FrontierMath tier 4 22.9% #5 / 38 #5 / 48 39.6% GPT-5.5 Pro
FrontierCode 38.5% #9 / 20 #28 / 43 54.6% Claude Opus 5.5
Hle 36.2% #6 / 18 #10 / 39 54.8% GPT-6 Astra
Gdp Pdf 21.0% #7 / 15 #22 / 37 35.0% GPT-6.1 Sol
Furniture Assembly 33.3% #9 / 14 #23 / 34 83.3% Claude Opus 5.5
Terminal-Bench 80.2% #3 / 31 (tied with 1) #3 / 33 84.7% GPT-5.5
SWE-bench Verified 83.5% #1 / 26 #1 / 32 83.5% Claude Opus 4.7
Blueprint Bench 2 24.5% #10 / 18 #23 / 31 54.4% Gemini 4 Argon
Gso 44.1% #4 / 17 #6 / 28 88.2% Claude Fable 5.1
Ebr Bench 19.0% #7 / 17 #17 / 27 76.2% GPT-6 Astra
Gbaeval 43.8% #10 / 20 #12 / 22 79.6% Claude Opus 5
Posttrainbench 28.6% #5 / 9 #7 / 11 41.8% Claude Fable 5
MirrorCode 31.1% #2 / 6 #5 / 9 77.4% Claude Opus 5.5
Osworld 2 18.2% #3 / 8 #4 / 9 31.4% Claude Opus 5
Exploitbench 26.5% #2 / 8 #2 / 8 47.4% GPT-5.5

For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.

Comparison

What to compare Claude Opus 4.7 with

The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.

ModelLabECIELOLiveBenchRelease
Qwen 3.8 Max Alibaba 156.88 1,536 78.5 Aug 2, 2026
Muse Spark 1.3 Meta AI 156.76 1,528 81.6 Sep 2, 2026
GPT-5.6 Luna OpenAI 156.33 1,478 73.6 Jul 9, 2026
Claude Opus 4.7 Anthropic 156.32 1,532 76.5 Apr 16, 2026
Grok 4.6 xAI 156.31 1,475 78.0 Aug 12, 2026
GPT-6 Luna OpenAI 156.31 n/a 72.0 Sep 22, 2026
GLM-5.3 Z.ai (Zhipu AI) 155.91 n/a 76.1 Aug 14, 2026
Head to head

Claude Opus 4.7 or Grok 4.6?

Grok 4.6 is the model from another lab closest to Claude Opus 4.7 in capability. Across the 25 tests both took, this is the score:

Claude Opus 4.7 ahead on 10 Grok 4.6 ahead on 14 tied on 1
Largest differences first
TestClaude Opus 4.7Grok 4.6
CritPt 12.0% 19.7%
Ebr Bench 19.0% 30.5%
Blueprint Bench 2 24.5% 33.2%
Apex Agents 49.2% 65.3%
FrontierCode 38.5% 48.0%
Chess puzzles 41.0% 33.0%
SimpleBench 61.7% 75.9%
Gdp Pdf 21.0% 17.2%
Game puzzles 28.0% 34.0%
Vending Bench 2 10,937 9,047

Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.

Cost per task

What Claude Opus 4.7 costs to deliver each task

Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.

ARC-AGI 2

The score scale does not start at zero.

0%22%43%65%86%US$ 0.10US$ 0.20US$ 0.50US$ 1US$ 2US$ 5cost per task (log scale)Claude Opus 4.7 · low · 62.1% · US$ 2.38lowClaude Opus 4.7 · medium · 67.5% · US$ 2.96mediumClaude Opus 4.7 · high · 68.3% · US$ 3.17highClaude Opus 4.7 · max · 75.8% · US$ 7.43maxGPT-5.6 Luna · low · 5.1% · US$ 0.07GPT-5.6 Luna · medium · 7.4% · US$ 0.11GPT-5.6 Luna · high · 29.3% · US$ 0.29GPT-5.6 Luna · xhigh · 47.6% · US$ 0.54GPT-5.6 Luna · max · 59.5% · US$ 0.67
Claude Opus 4.7GPT-5.6 Luna

ARC-AGI

The score scale does not start at zero.

25%44%63%81%100%US$ 0.05US$ 0.10US$ 0.20US$ 0.50US$ 1US$ 2cost per task (log scale)Claude Opus 4.7 · low · 91.0% · US$ 0.76lowClaude Opus 4.7 · medium · 91.0% · US$ 1.04mediumClaude Opus 4.7 · high · 93.5% · US$ 1.41highClaude Opus 4.7 · max · 92.0% · US$ 2.58maxGPT-5.6 Luna · low · 34.2% · US$ 0.04GPT-5.6 Luna · medium · 56.5% · US$ 0.06GPT-5.6 Luna · high · 76.5% · US$ 0.13GPT-5.6 Luna · xhigh · 87.7% · US$ 0.21GPT-5.6 Luna · max · 88.0% · US$ 0.32
Claude Opus 4.7GPT-5.6 Luna
Family

Other models from Anthropic

The models from the same lab closest in time, newest first.

ModelReleaseECIELOLiveBenchAccess
Claude Fable 5.1 Sep 1, 2026 164.65 1,586 83.4 API access
Claude Opus 5 Jul 24, 2026 162.60 1,573 80.1 API access
Claude Sonnet 5 Jun 30, 2026 156.36 1,491 76.0 API access
Claude Fable 5 Jun 9, 2026 162.13 1,542 83.0 API access
Claude Opus 4.8 May 28, 2026 158.26 1,507 76.2 API access
Claude Opus 4.7 Apr 16, 2026 156.32 1,532 76.5 API access
Claude Sonnet 4.6 Feb 17, 2026 152.39 1,504 73.0 API access
Claude Opus 4.6 Feb 5, 2026 155.27 1,550 74.5 API access
Claude Opus 4.5 Nov 24, 2025 150.17 1,504 72.6 API access
Your rating

Have you used Claude Opus 4.7? Rate it

n/a out of 5
no votes yet

Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.

Tap a star to vote.
Frequently asked questions

What people ask about Claude Opus 4.7

Does Claude Opus 4.7 still work?

Yes. It is a legacy model, but it is still available on the Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic guarantees it will not be retired before April 16, 2027 on the platforms it operates, with at least 60 days' notice.

How much does Claude Opus 4.7 cost?

US$ 5 per million input tokens and US$ 25 per million output, the same as Opus 4.6. Cache reads cost US$ 0.50, cache writes US$ 6.25 (5 minutes) or US$ 10 (1 hour), and batch is US$ 2.50 and US$ 12.50. It has no fast mode.

Which model replaced Opus 4.7?

Opus 4.8, released on May 28, 2026 at the same price. Then came Opus 5, on July 24, and Opus 5.5, on September 22, 2026, which is where Anthropic's documentation now says to migrate.

Is it worth migrating from Opus 4.7 to Opus 5.5?

For most uses, yes: Opus 5.5 costs US$ 4 and US$ 20, rereads the cache for US$ 0.20 and is Anthropic's recommended starting point. Check three things first: thinking becomes always on, forced tool use stops working and Priority Tier does not exist on Opus 5.5.

Where does Claude Opus 4.7 rank among AI models?

In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Opus 4.7 ranks #26 out of 267 models, with an ECI of 156.32. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.

Is Claude Opus 4.7 better than Grok 4.6?

Across the 25 tests both models took and Epoch AI compiled, Claude Opus 4.7 comes out ahead on 10, behind on 14 and tied on 1. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.

What is Claude Opus 4.7 best at?

Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Claude Opus 4.7 are: Vending Bench 2 (#1 of 43); SWE-bench Verified (#1 of 26); WebDev Arena (#4 of 55). The yardstick is its position among the models of that period that took the same test.

Where do the numbers on this page come from?

From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.

Sources

Where each number comes from

  1. Anthropic, Claude Opus 4.7 announcement
  2. Anthropic, Claude Opus 4.7 system card
  3. Anthropic, Claude Opus 4.7 overview (documentation)
  4. Anthropic, pricing (documentation)
  5. Anthropic, model deprecations (documentation)
  6. Anthropic, migrating to Claude Opus 5.5 (documentation)
  7. Anthropic, effort levels (documentation)
  8. Anthropic, prompt caching (documentation)
  9. Anthropic, platform release notes (documentation)
  10. Anthropic, Claude Opus 4.8 announcement (comparison with Opus 4.7)
  11. Anthropic, Claude Opus 4.8 system card

Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.

The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.

LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.

← See the full ranking