Home › AI Benchmark › Claude Opus 5
Anthropic released on Jul 24, 2026 API access

Claude Opus 5

Claude Opus 5 was Anthropic's main Opus model from Jul 24, 2026 until Sep 22, 2026, when Opus 5.5 arrived. It is now a legacy model: still available, and Anthropic guarantees it will not be retired before Jul 24, 2027, but it costs US$ 5 per million input tokens and US$ 25 per million output, against US$ 4 and US$ 20 for its successor, and rereads the cache for US$ 0.50 instead of US$ 0.20. In the Opus 5.5 launch table, it loses all nine tests. It is only worth keeping while you adapt what 5.5 rejects: thinking turned off, forced tool use and the old computer use tool.

#7 of 267 in the ECI ranking
162.60 ECI, the Epoch AI capability index
1,573 Arena ELO, human preference · #4
80.1 LiveBench, overall average from 0 to 100 · #10
No votes yet. Rate it.
Spec sheet

Claude Opus 5 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-5
Release
Jul 24, 2026
Status
legacy; the successor is Opus 5.5, released on Sep 22, 2026
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
US$ 10 and US$ 50, Anthropic API only
Context window
1 million tokens, no surcharge
Max output
128 thousand tokens (300 thousand in batch, beta)
Reliable knowledge
through May 2026
Reasoning
adaptive, on by default; can be turned off only at high effort or below; default effort high
Where it runs
Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS
Data retention
available with zero data retention
Retirement
not before Jul 24, 2027
Analysis

Where Claude Opus 5 does well, where it falls behind and who it is for

Where it does well

  • Terminal coding: 43.3% on Frontier-Bench v0.1 in the launch table, against 34.4% for GPT-5.6 Sol, 33.7% for Fable 5 and 21.1% for Opus 4.8.
  • Office work: 1,861 points on GDPval-AA v2, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736), at half Fable 5's price per token.
  • Novel problems: 30.2% on ARC-AGI-3, against 7.8% for GPT-5.6 Sol and 1.5% for Opus 4.8, in Anthropic's figures.
  • Behavior: 2.30 for misaligned behavior on Anthropic's automated audit (lower is better), against 2.85 for Opus 4.8 and 3.35 for Sonnet 5.
  • No retention requirement: runs with zero data retention, unlike Fable 5, and offers a fast mode of up to 2.5 times the speed.

Where it falls behind

  • Overtaken by its successor: in the Opus 5.5 launch table, it loses all nine tests; on Terminal-Bench 4.0 it scores 52.3% against 66.4%.
  • Pricier cache: cache reads cost US$ 0.50 per million; on Opus 5.5, US$ 0.20, or 60% less. In agent and coding work, Anthropic says this is the line that weighs most on the bill.
  • Hallucinates slightly more: in the system card, it is 11% more accurate than Opus 4.8 on AA-Omniscience, but hallucinates 6% more.
  • Talks too much: the documentation warns that responses run longer than earlier Opus models, that it delegates to subagents more readily and that lowering effort does not shorten the text.
  • It did not lead everything even at launch: 68.8% on DeepSWE v1.1, against 72.7% for GPT-5.6 Sol and 69.7% for Fable 5; on FrontierCode, 53.4% against 53.5% for Fable 5.

Who it makes sense for

  • Integrations that run with thinking turned off (at high effort or below): on Opus 5.5, the same request returns a 400 error.
  • Flows that force a tool with tool_choice any or tool, accepted on Opus 5 and rejected on Opus 5.5.
  • Computer use agents on the Anthropic API or Google Cloud that still use the computer_20251124 tool and have not moved to the new toolset.
  • Systems already validated on Opus 5 that need time to rerun their tests: retirement will not come before Jul 24, 2027, with at least 60 days' notice.

When to pick another model

  • New projects: Anthropic's documentation says to start with Opus 5.5, which costs US$ 4 and US$ 20 per million tokens.
  • Apps that reread a lot of cached context: Opus 5.5 charges US$ 0.20 for cache reads, 60% less.
  • Running at default effort: on Terminal-Bench 4.0, Opus 5 at high scored 47.0% for US$ 10.64 per task; Opus 5.5 at medium, its default, scored 57.6% for US$ 2.94.
  • Live chat with short answers: responses run long by default, and Sonnet 5.5 costs US$ 2 and US$ 10.

What Anthropic published at launch

Opus 5 shipped on Jul 24, 2026 with the pitch of coming close to Fable 5 at half the price. It kept Opus 4.8's price and became the default model on the Max plan. In Anthropic's figures, it more than doubled Opus 4.8 on Frontier-Bench v0.1 at a lower cost per task.

In the table below, from the announcement and the system card (max effort, averaged over 5 trials), Opus 5 leads eight of the twelve rows. Compare with care: in the Fable 5.1 announcement, Anthropic itself gave it 1,824 on GDPval-AA v2, not 1,861.

Opus 5 in the figures Anthropic published at launch
Opus 5
Anthropic		Fable 5
Anthropic		Opus 4.8
Anthropic		GPT-5.6 Sol
OpenAI
Frontier-Bench v0.1
terminal coding		 43.3%33.7%21.1%34.4%
GDPval-AA v2
office work, Elo		 1,8611,7471,5931,736
ARC-AGI-3
novel problems		 30.2%not published1.5%7.8%
BrowseComp
agentic search		 90.8%87.4%84.3%90.4%
Humanity's Last Exam
without tools		 56.3%56.5%49.8%not published
Humanity's Last Exam
with tools		 64.7%63.9%57.9%not published
OSWorld 2.0
computer use		 70.6%66.1%55.7%62.6%
DeepSWE v1.1
agentic coding		 68.8%69.7%59.0%72.7%
FrontierCode v1.1, Main
coding		 53.4%53.5%46.5%47.5%
AutomationBench
business workflows		 26.0%17.4%17.0%18.1%
Legal Agent Benchmark
legal, held-out set		 11.7%13.3%10.4%2.5%
BioMysteryBench
biology, hard set		 49.4%46.5%42.4%not published

Figures from the Claude Opus 5 announcement and system card, table 8.1.A. The dot marks the best in each row. On Frontier-Bench, the system card gives Fable 5 33.8; the announcement, 33.7%.

Is Opus 5 still worth using in 2026?

For new work, no. Anthropic classifies Opus 5 as legacy and recommends Opus 5.5, which costs 20% less per token, rereads the cache for 60% less and, according to the company, runs 40% cheaper on typical workloads.

As a bridge, yes. Anthropic guarantees it will not be retired before Jul 24, 2027 on the platforms it operates (the API, Claude Platform on AWS and Microsoft Foundry) and gives at least 60 days' notice by email; Amazon Bedrock and Google Cloud set their own dates. After retirement, requests fail.

Opus 5 against current models, price per million tokens
Opus 5
legacy		Opus 5.5
current		Sonnet 5.5
current		Fable 5.1
current
Input US$ 5US$ 4US$ 2US$ 10
Output US$ 25US$ 20US$ 10US$ 50
Cache read US$ 0.50US$ 0.20US$ 0.20US$ 0.25
Cache write
5 minutes		 US$ 6.25US$ 5US$ 2.50US$ 12.50
Fast mode US$ 10 and US$ 50US$ 8 and US$ 40not availablenot available
Default effort highmediumhighhigh
Reliable knowledge May 2026June 2026June 2026June 2026
Retirement
not before		 Jul 24, 2027Sep 22, 2027Sep 28, 2027Sep 1, 2027

Anthropic documentation, pricing, models overview and model deprecations, checked on Oct 1, 2026.

What changes when you migrate to Opus 5.5

Anthropic's migration guide lists what breaks when you swap claude-opus-5 for claude-opus-5-5. Nothing requires rewriting the product, but some items fail the call with a 400 error:

  • Thinking can no longer be turned off: thinking set to disabled or a manual budget returns a 400 error. Effort becomes the only control, and its default drops from high to medium; rerun your effort sweep and, at xhigh or max, start max_tokens at 64 thousand.
  • tool_choice any and tool return a 400 error; use auto with strict tools or structured outputs.
  • On the Anthropic API and Google Cloud, computer use requires computer_toolset_20260801; computer_20251124 is only still accepted on Amazon Bedrock.
  • Text the model writes between tool calls now arrives in thinking blocks, empty by default. If you show that progress to users, request display updates (beta) or summarized.
  • There are more refusal categories: besides cyber, bio and reasoning_extraction.
Default effort against default effort, in Anthropic's charts
Opus 5
high		Opus 5.5
medium
Terminal-Bench 4.0
score		 47.0%57.6%
Terminal-Bench 4.0
cost per task		 US$ 10.64US$ 2.94
CursorBench 4.0
score		 44.7%52.5%
CursorBench 4.0
cost per task		 US$ 9US$ 2.90
GDPval-AA v2.1
Elo		 1,5811,576
GDPval-AA v2.1
cost per task		 US$ 3.03US$ 0.86

Per-effort points from the charts in the Claude Opus 5.5 announcement (Sep 22, 2026), with cost per task as measured by Anthropic. The dot marks the best in each row.

In practice

Examples of using Claude Opus 5

An old integration with thinking turned off

A ticket classifier calls Opus 5 with thinking turned off to answer quickly, and it keeps working until retirement. Before moving to Opus 5.5, where that call returns a 400 error, the team tests low effort as a replacement.

Code review while the migration is pending

According to the documentation, Opus 5 finds real bugs at a high rate per pass, even at lower effort. You can keep code review on it while the successor is validated; ask it to report everything and filter afterwards, because it follows requests like "only report high-severity issues" literally.

A side-by-side test before switching off

Pick ten real tasks and run them on Opus 5 at high and Opus 5.5 at medium. Compare accuracy and cost per task, not price per token. In Anthropic's figures, GDPval-AA stayed almost the same (1,581 against 1,576) while cost fell from US$ 3.03 to US$ 0.86.

What the tests say

Where Claude Opus 5 stands out and where it falls behind, by the numbers

Claude Opus 5 has published results on 31 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Oct 27, 2025 and Oct 22, 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.

Gbaeval79.6%
#1 of 22 in its generation · #1 of 22 overall
Osworld 231.4%
#1 of 9 in its generation · #1 of 9 overall
SimpleBench80.6%
#2 of 38 in its generation · #2 of 74 overall
Dtbench97.6%
#3 of 69 in its generation, tied with 1 · #3 of 148 overall
Lmca63.30
#4 of 69 in its generation · #4 of 121 overall

Where it falls behind

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

Blueprint Bench 230.4%
#17 of 30 in its generation · record: 54.4% (Gemini 4 Argon)
See all 31 tests, one by one
TestScoreIn its generationOverallTest record
GPQA Diamond 93.9% #14 / 81 #14 / 173 96.4% Gemini 4 Argon
AIME 2024 and 2025 (OTIS) 98.9% #17 / 83 (tied with 3) #17 / 168 100.0% GPT-5.5
Dtbench 97.6% #3 / 69 (tied with 1) #3 / 148 98.9% Claude Opus 5.5
Lmca 63.30 #4 / 69 #4 / 121 68.23 Claude Opus 5.5
CritPt 29.1% #11 / 79 #11 / 112 32.3% GPT-5.6 Sol
WeirdML 91.8% #4 / 54 #4 / 112 93.6% GPT-6 Astra
SciCode 56.4% #23 / 76 #23 / 104 66.9% Claude Opus 5.5
Frontiermath Tiers 1 3 V2 85.6% #12 / 68 #12 / 86 93.7% GPT-6 Astra
ARC-AGI 2 90.4% #5 / 50 #5 / 83 95.0% GPT-6 Astra
WebDev Arena 1,687 #7 / 75 #7 / 82 1,820 Claude Opus 5.5
ARC-AGI 97.5% #7 / 50 (tied with 2) #7 / 81 98.5% Claude Fable 5
SimpleQA Verified 59.9% #17 / 64 #17 / 81 76.0% Gemini 4 Argon
Chess puzzles 43.0% #28 / 64 (tied with 1) #29 / 80 76.0% GPT-6 Astra
Game puzzles 59.0% #6 / 62 #6 / 77 84.0% GPT-6 Astra
ProofBench 99.0% #4 / 70 (tied with 3) #4 / 74 100.0% Claude Fable 5.1
SimpleBench 80.6% #2 / 38 #2 / 74 81.9% Claude Fable 5
Frontiermath Tier 4 V2 73.2% #11 / 58 #11 / 67 100.0% GPT-6.1 Sol
Vending Bench 2 11,182 #4 / 49 #4 / 57 15,515 GPT-6 Astra
Apex Agents 65.8% #6 / 49 #6 / 50 82.4% Gemini 4 Argon
FrontierCode 53.4% #3 / 43 #3 / 43 54.6% Claude Opus 5.5
Gdp Pdf 24.0% #14 / 37 (tied with 1) #14 / 37 35.0% GPT-6.1 Sol
Furniture Assembly 60.8% #7 / 34 #7 / 34 83.3% Claude Opus 5.5
BALROG 63.4% #2 / 9 #2 / 32 68.3% GPT-6 Astra
Blueprint Bench 2 30.4% #17 / 30 #17 / 31 54.4% Gemini 4 Argon
Ebr Bench 45.7% #7 / 24 #7 / 27 76.2% GPT-6 Astra
DeepSWE 73.6% #3 / 26 #3 / 26 74.1% GPT-6 Astra
Gbaeval 79.6% #1 / 22 #1 / 22 79.6% Claude Opus 5
FrontierSWE 52.0% #6 / 20 #6 / 20 65.5% GPT-6 Astra
CursorBench 46.6% #6 / 16 #6 / 16 57.8% Claude Opus 5.5
Posttrainbench 35.0% #3 / 11 #3 / 11 41.8% Claude Fable 5
Osworld 2 31.4% #1 / 9 #1 / 9 31.4% Claude Opus 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 5 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
Gemini 4 Argon Google DeepMind 164.56 1,608 n/a Sep 30, 2026
Claude Opus 5 Anthropic 162.60 1,573 80.1 Jul 24, 2026
GPT-6 Sol OpenAI 162.55 n/a 79.3 Sep 22, 2026
GPT-5.5 Pro OpenAI 161.97 n/a n/a Apr 23, 2026
Kimi K3 Moonshot 157.42 1,528 79.2 Jul 16, 2026
Gemini 3.7 Flash Google DeepMind 157.22 1,551 78.8 Aug 13, 2026
Qwen 3.8 Max Alibaba 156.88 1,536 78.5 Aug 2, 2026
Head to head

Claude Opus 5 or GPT-6 Sol?

GPT-6 Sol is the model from another lab closest to Claude Opus 5 in capability. Across the 22 tests both took, this is the score:

Claude Opus 5 ahead on 9 GPT-6 Sol ahead on 13
Largest differences first
TestClaude Opus 5GPT-6 Sol
Chess puzzles 43.0% 56.0%
Vending Bench 2 11,182 14,428
Frontiermath Tier 4 V2 73.2% 90.0%
Blueprint Bench 2 30.4% 36.9%
Apex Agents 65.8% 54.3%
ProofBench 99.0% 83.0%
Ebr Bench 45.7% 53.3%
Gdp Pdf 24.0% 26.4%
FrontierCode 53.4% 49.3%
Lmca 63.30 59.08

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 5 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.

CursorBench

The score scale does not start at zero.

39%43%47%51%55%US$ 4US$ 6US$ 9US$ 12cost per taskClaude Opus 5 · low · 40.7% · US$ 4.87lowClaude Opus 5 · medium · 43.3% · US$ 6.94mediumClaude Opus 5 · high · 44.7% · US$ 9.00highClaude Opus 5 · xhigh · 46.1% · US$ 11.43xhighClaude Opus 5 · max · 46.6% · US$ 11.95maxGemini 4 Argon · minimal · 48.8% · US$ 3.54Gemini 4 Argon · low · 50.6% · US$ 4.09Gemini 4 Argon · medium · 52.8% · US$ 4.39Gemini 4 Argon · high · 52.9% · US$ 4.57
Claude Opus 5Gemini 4 Argon

DeepSWE

The score scale does not start at zero.

51%57%64%70%77%US$ 2US$ 5US$ 8US$ 12cost per taskClaude Opus 5 · low · 58.1% · US$ 1.66lowClaude Opus 5 · medium · 68.9% · US$ 3.29mediumClaude Opus 5 · high · 72.8% · US$ 6.08highClaude Opus 5 · xhigh · 73.2% · US$ 9.07xhighClaude Opus 5 · max · 73.6% · US$ 11.84maxGemini 3.7 Flash · low · 53.8% · US$ 1.83Gemini 3.7 Flash · medium · 65.5% · US$ 2.03Gemini 3.7 Flash · high · 65.3% · US$ 2.18
Claude Opus 5Gemini 3.7 Flash
Family

Other models from Anthropic

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

ModelReleaseECIELOLiveBenchAccess
Claude Haiku 5.5 Oct 7, 2026 158.98 n/a 72.1 API access
Claude Sonnet 5.5 Sep 28, 2026 164.45 1,537 77.8 API access
Claude Opus 5.5 Sep 22, 2026 166.82 1,584 83.2 API access
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
Your rating

Have you used Claude Opus 5? 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 5

Does Claude Opus 5 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 July 24, 2027 on the platforms it operates and gives at least 60 days' notice. Bedrock and Google Cloud set their own schedules. After that date, requests to claude-opus-5 fail.

How much does Claude Opus 5 cost?

US$ 5 per million input tokens and US$ 25 per million output. 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. Fast mode, on the Anthropic API only, costs US$ 10 and US$ 50. The price has not changed since launch.

What is the difference between Opus 5 and Opus 5.5?

Opus 5.5 costs US$ 4 and US$ 20, rereads the cache for US$ 0.20 and beats Opus 5 on all nine tests in its launch table, such as Terminal-Bench 4.0 (66.4% against 52.3%). In exchange, it does not let you turn thinking off, does not accept forced tool use and defaults to medium effort instead of high.

Opus 5 or Fable 5?

In the Opus 5 launch table, it beats Fable 5 on terminal coding (43.3% against 33.7%) and office work (1,861 against 1,747) at half the price per token. Fable 5 wins on FrontierCode and the Legal Agent Benchmark. Both are legacy today: their successors are Opus 5.5 and Fable 5.1.

Where does Claude Opus 5 rank among AI models?

In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Opus 5 ranks #7 out of 267 models, with an ECI of 162.60. 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 5 better than GPT-6 Sol?

Across the 22 tests both models took and Epoch AI compiled, Claude Opus 5 comes out ahead on 9, behind on 13 and tied on 0. 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 5 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 5 are: Gbaeval (#1 of 22); Osworld 2 (#1 of 9); SimpleBench (#2 of 38). 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 5 announcement
  2. Anthropic, Claude Opus 5 system card
  3. Anthropic, Claude Opus 5 overview (documentation)
  4. Anthropic, pricing (documentation)
  5. Anthropic, model deprecations (documentation)
  6. Anthropic, migrating to Claude Opus 5.5 (documentation)
  7. Anthropic, what's new in Claude Opus 5.5 (documentation)
  8. Anthropic, prompting Claude Opus 5 (documentation)
  9. Anthropic, models overview (documentation)
  10. Anthropic, Claude Opus 5.5 announcement (comparison with Opus 5)
  11. Anthropic, Claude Fable 5.1 announcement (later figures for Opus 5)

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