Home › AI Benchmark › Claude Opus 4.8
Anthropic released on May 28, 2026 API access

Claude Opus 4.8

Claude Opus 4.8 shipped on May 28, 2026, presented by Anthropic as its most capable model, and twelve days later Fable 5 was already above it; Opus 5 followed on Jul 24. 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 May 28, 2027. It still has a role no other model has: it is the model that answers when Fable 5, Fable 5.1, Opus 5 and Opus 5.5 block a cybersecurity request.

#16 of 267 in the ECI ranking
158.26 ECI, the Epoch AI capability index
1,507 Arena ELO, human preference · #49
76.2 LiveBench, overall average from 0 to 100 · #24
No votes yet. Rate it.
Spec sheet

Claude Opus 4.8 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-4-8
Release
May 28, 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
US$ 10 and US$ 50, up to 2.5 times faster, Anthropic API only
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
1,024 tokens (512 on current models)
Where it runs
Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS
Retention and capacity
available with zero data retention; supports Priority Tier
Retirement
not before May 28, 2027
Analysis

Where Claude Opus 4.8 does well, where it falls behind and who it is for

Where it does well

  • Coding: 69.2% on SWE-Bench Pro at launch, against 64.3% for Opus 4.7, 58.6% for GPT-5.5 and 54.2% for Gemini 3.1 Pro.
  • Office work: 1,890 on GDPval-AA, against 1,769 for GPT-5.5 and 1,753 for Opus 4.7.
  • Financial analysis: 53.9% on Finance Agent v2, ahead of GPT-5.5 (51.8%) and Opus 4.7 (51.5%).
  • Honesty: according to Anthropic, around four times less likely than Opus 4.7 to let flaws in code it wrote pass unremarked.
  • Predictable in the API: it follows prompts literally, especially at low effort, and only thinks when the request turns thinking on, which helps structured extraction and pipelines.

Where it falls behind

  • Terminal: 74.6% on Terminal-Bench 2.1, behind GPT-5.5's 78.2% in its own launch table.
  • Previous generation: at the Opus 5 launch, it scored 21.1% on Frontier-Bench v0.1 against 43.3%, and 1.5% on ARC-AGI-3 against 30.2%, at the same price per token.
  • Prompt injection: the system card says it is somewhat less robust than Opus 4.7 in several agentic contexts; the safeguards close the gap in practice.
  • Effort-sensitive: Anthropic says effort matters more for it than for any prior Opus and that, at low, there is a risk of under-thinking on moderately complex tasks.
  • Cache: it only caches prompts from 1,024 tokens, against 512 on current models, and rereads for US$ 0.50, against US$ 0.20 on Opus 5.5.

Who it makes sense for

  • Fallback for refusals: it is where newer Fable and Opus models send blocked cybersecurity requests, billed at its own rates.
  • Accounts with a Priority Tier contract, which Opus 4.8 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 4.8 and breaks on Opus 5.5.
  • Systems validated on it that need time to rerun tests: it will not be retired before May 28, 2027, with at least 60 days' notice.

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.
  • Long agent work: at the same price, Opus 5 more than doubled its score on Frontier-Bench v0.1 at a lower cost per task.
  • Fast mode: Opus 5.5's costs US$ 8 and US$ 40, against US$ 10 and US$ 50.
  • Tasks that need recent knowledge: its reliable cutoff is January 2026; Opus 5.5's is June 2026.

What Anthropic published at launch

Opus 4.8 arrived on May 28, 2026 as an update to Opus 4.7 at the same price, and Anthropic itself called it a modest but tangible improvement. It came with effort control in Claude.ai, dynamic workflows in Claude Code, with hundreds of parallel subagents, and a fast mode three times cheaper than on previous models.

Two caveats from Anthropic itself: on Terminal-Bench 2.1, all models ran on the public Terminus-2 harness, and GPT-5.5 scores 83.4% with the Codex CLI; on OSWorld-Verified, the table shows 82.8% for Opus 4.7, but the footnote says 82.3%.

Opus 4.8 in the figures Anthropic published at launch
Opus 4.8
Anthropic		Opus 4.7
Anthropic		GPT-5.5
OpenAI		Gemini 3.1 Pro
Google
SWE-Bench Pro
coding		 69.2%64.3%58.6%54.2%
Terminal-Bench 2.1
terminal coding		 74.6%66.1%78.2%70.3%
Humanity's Last Exam
without tools		 49.8%46.9%41.4%44.4%
Humanity's Last Exam
with tools		 57.9%54.7%52.2%51.4%
OSWorld-Verified
computer use		 83.4%82.8%78.7%76.2%
GDPval-AA
office work, Elo		 1,8901,7531,7691,314
Finance Agent v2
financial analysis		 53.9%51.5%51.8%43.0%

Figures from the Claude Opus 4.8 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.

Anthropic's fallback model

Opus 4.8 picked up a second job. When Fable 5's classifiers block a cybersecurity request, the answer comes from Opus 4.8, and the user is told. Opus 5 does the same in Claude.ai, Claude Code and Cowork, and Opus 5.5 reroutes most cybersecurity tasks to it. In Fable 5.1's tests, blocked cyber tasks were also completed by Opus 4.8.

On the API, the fallback is optional: the fallbacks parameter (beta) reruns a refused request on another model, and the documentation's example uses claude-opus-4-8. That part is billed at the fallback model's rates, so if you configure it you may see Opus 4.8 on your bill even after migrating.

Is Opus 4.8 still worth using in 2026?

As your main model, no. At the same price per token, Opus 5 already beat it comfortably at its own launch, and Opus 5.5 costs less than both. The documentation says to migrate straight to Opus 5.5.

There are three reasons to keep it longer: the fallback role, Priority Tier contracts and integrations that depend on running without thinking or forcing a tool. Anthropic guarantees it will not be retired before May 28, 2027 on the platforms it operates and gives at least 60 days' notice; Amazon Bedrock and Google Cloud set their own dates.

Opus 4.8 against its successors, price per million tokens
Opus 4.8
legacy		Opus 5
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 US$ 10 and US$ 50US$ 10 and US$ 50US$ 8 and US$ 40
Reasoning
when not requested		 offonalways on
Default effort highhighmedium
Cache minimum 1,024 tokens512 tokens512 tokens
Reliable knowledge January 2026May 2026June 2026
Retirement
not before		 May 28, 2027Jul 24, 2027Sep 22, 2027

Anthropic documentation: pricing, each model's page, migration guide and model deprecations, checked on Oct 1, 2026.

What changes when you migrate to Opus 5.5

The migration guide treats Opus 4.8 almost like Opus 5: both accept thinking turned off, forced tool use and the old computer use tool, and both default to high effort. All of that changes on Opus 5.5:

  • Thinking by default: on Opus 4.8, a request without a thinking field runs without thinking; on Opus 5.5, it always thinks. Thinking tokens are billed as output, so the same work can produce more output.
  • Turning thinking off or setting a manual budget returns a 400 error; the control is effort, which defaults to medium. 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.
  • Caching starts at 512 tokens, with no code change, and Priority Tier goes away.
In practice

Examples of using Claude Opus 4.8

A fallback set up for refusals

A security tool runs on Opus 5.5 with fallbacks in default mode. When the classifier blocks a cybersecurity request, the API reruns it on Opus 4.8, and the response names the model that served it, billed at its rates.

Structured extraction without thinking

A pipeline reads contracts and returns fields through forced tool use, without thinking, to stay fast and predictable. It works on Opus 4.8. To move to Opus 5.5, you need to replace forced tool use with auto and strict tools and accept thinking, controlled by effort.

A test before migrating

Pick ten real tasks and run them on Opus 4.8 at xhigh, Anthropic's recommendation for coding, and on Opus 5.5 at medium. Compare accuracy, output tokens and cost per task, because always-on thinking changes the bill.

What the tests say

Where Claude Opus 4.8 stands out and where it falls behind, by the numbers

Claude Opus 4.8 has published results on 38 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Aug 31, 2025 and Aug 26, 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.

Lmca57.51
#4 of 65 in its generation · #8 of 121 overall
FrontierMath tier 431.3%
#3 of 33 in its generation · #3 of 48 overall
WeirdML82.9%
#5 of 57 in its generation · #8 of 112 overall
Gbaeval70.9%
#3 of 22 in its generation · #3 of 22 overall
ALE-Bench1,564
#7 of 56 in its generation · #13 of 82 overall

Where it falls behind

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

ForecastBench58.60
#21 of 25 in its generation · record: 62.50 (o3)
Blueprint Bench 214.5%
#16 of 22 in its generation · record: 54.4% (Gemini 4 Argon)
Btf314.0%
#5 of 7 in its generation · record: 14.9% (GLM-5.3)
See all 38 tests, one by one
TestScoreIn its generationOverallTest record
GPQA Diamond 91.0% #21 / 72 (tied with 1) #28 / 173 96.4% Gemini 4 Argon
AIME 2024 and 2025 (OTIS) 98.3% #10 / 72 (tied with 1) #22 / 168 100.0% GPT-5.5
Dtbench 94.9% #11 / 67 #18 / 148 98.9% Claude Opus 5.5
Lmca 57.51 #4 / 65 #8 / 121 68.23 Claude Opus 5.5
CritPt 20.9% #10 / 65 (tied with 1) #21 / 112 32.3% GPT-5.6 Sol
WeirdML 82.9% #5 / 57 #8 / 112 93.6% GPT-6 Astra
SciCode 53.5% #19 / 61 (tied with 1) #35 / 104 66.9% Claude Opus 5.5
Frontiermath Tiers 1 3 V2 80.0% #9 / 55 #17 / 86 93.7% GPT-6 Astra
ARC-AGI 2 72.1% #12 / 46 (tied with 1) #19 / 83 95.0% GPT-6 Astra
ALE-Bench 1,564 #7 / 56 #13 / 82 2,951 GPT-6 Astra
WebDev Arena 1,559 #14 / 65 #28 / 82 1,820 Claude Opus 5.5
ARC-AGI 92.5% #14 / 46 (tied with 1) #21 / 81 98.5% Claude Fable 5
SimpleQA Verified 53.0% #13 / 53 #22 / 81 76.0% Gemini 4 Argon
Chess puzzles 52.0% #7 / 53 #15 / 80 76.0% GPT-6 Astra
Game puzzles 36.0% #9 / 49 #19 / 77 84.0% GPT-6 Astra
ProofBench 69.0% #7 / 52 #17 / 74 100.0% Claude Fable 5.1
SimpleBench 64.8% #15 / 41 #15 / 74 81.9% Claude Fable 5
Frontiermath Tier 4 V2 56.1% #9 / 44 #17 / 67 100.0% GPT-6.1 Sol
FrontierMath 47.2% #5 / 33 #5 / 65 51.7% GPT-5.5
ForecastBench 58.60 #21 / 25 #39 / 62 62.50 o3
Vending Bench 2 5,787 #16 / 44 #22 / 57 15,515 GPT-6 Astra
Apex Agents 48.9% #16 / 35 #28 / 50 82.4% Gemini 4 Argon
FrontierMath tier 4 31.3% #3 / 33 #3 / 48 39.6% GPT-5.5 Pro
FrontierCode 46.5% #5 / 29 #14 / 43 54.6% Claude Opus 5.5
Gdp Pdf 24.0% #5 / 24 (tied with 1) #14 / 37 35.0% GPT-6.1 Sol
EnigmaEval 23.5% #6 / 17 #6 / 36 39.3% Claude Fable 5
Furniture Assembly 42.5% #5 / 20 (tied with 1) #15 / 34 83.3% Claude Opus 5.5
Blueprint Bench 2 14.5% #16 / 22 #25 / 31 54.4% Gemini 4 Argon
Gso 47.1% #3 / 16 #5 / 28 88.2% Claude Fable 5.1
Ebr Bench 28.6% #6 / 17 #14 / 27 76.2% GPT-6 Astra
DeepSWE 59.0% #12 / 24 #14 / 26 74.1% GPT-6 Astra
DeepResearch Bench 50.2% #6 / 15 #6 / 24 55.3% Claude Opus 4.6
Surface Evolver Bench 87.5% #5 / 20 #5 / 24 95.0% Claude Fable 5
Gbaeval 70.9% #3 / 22 #3 / 22 79.6% Claude Opus 5
Rli 8.3% #2 / 9 #5 / 15 21.3% Claude Opus 5.5
Posttrainbench 33.8% #4 / 11 #4 / 11 41.8% Claude Fable 5
Osworld 2 20.6% #3 / 9 #3 / 9 31.4% Claude Opus 5
Btf3 14.0% #5 / 7 #6 / 8 14.9% GLM-5.3

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 4.8 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
GPT-5.4 Pro OpenAI 159.00 n/a n/a Mar 5, 2026
GPT-5.5 OpenAI 158.95 1,533 80.2 Apr 23, 2026
Claude Opus 4.8 Anthropic 158.26 1,507 76.2 May 28, 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
Gemini 3.8 Flash Google DeepMind 157.20 1,551 75.8 Sep 2, 2026
Qwen 3.8 Max Alibaba 156.88 1,536 78.5 Aug 2, 2026
Head to head

Claude Opus 4.8 or GPT-5.5?

GPT-5.5 is the model from another lab closest to Claude Opus 4.8 in capability. Across the 37 tests both took, this is the score:

Claude Opus 4.8 ahead on 9 GPT-5.5 ahead on 28
Largest differences first
TestClaude Opus 4.8GPT-5.5
Blueprint Bench 2 14.5% 36.2%
Osworld 2 20.6% 13.0%
Game puzzles 36.0% 56.0%
ProofBench 69.0% 50.0%
Rli 8.3% 6.3%
Gbaeval 70.9% 53.2%
CritPt 20.9% 27.1%
Vending Bench 2 5,787 7,524
Frontiermath Tier 4 V2 56.1% 72.5%
Posttrainbench 33.8% 27.2%

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 4.8 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.

1%25%49%72%96%US$ 0.50US$ 1US$ 2cost per task (log scale)Claude Opus 4.8 · low · 62.2% · US$ 1.68lowClaude Opus 4.8 · medium · 71.7% · US$ 2.39mediumClaude Opus 4.8 · high · 72.1% · US$ 2.74highGPT-5.5 · low · 33.3% · US$ 0.35GPT-5.5 · medium · 70.4% · US$ 0.86GPT-5.5 · high · 83.3% · US$ 1.45GPT-5.5 · xhigh · 85.0% · US$ 1.87Kimi K3 · low · 12.4% · US$ 0.25Kimi K3 · high · 55.0% · US$ 0.95Kimi K3 · max · 60.4% · US$ 1.59
Claude Opus 4.8GPT-5.5Kimi K3

ARC-AGI

The score scale does not start at zero.

61%71%80%90%99%US$ 0.20US$ 0.50US$ 1US$ 2cost per task (log scale)Claude Opus 4.8 · low · 88.0% · US$ 0.67lowClaude Opus 4.8 · medium · 91.5% · US$ 0.91mediumClaude Opus 4.8 · high · 92.0% · US$ 1.04highClaude Opus 4.8 · max · 92.5% · US$ 2.33maxGPT-5.5 · low · 76.2% · US$ 0.20GPT-5.5 · medium · 92.2% · US$ 0.39GPT-5.5 · high · 94.5% · US$ 0.56GPT-5.5 · xhigh · 95.0% · US$ 0.73Kimi K3 · low · 65.7% · US$ 0.18Kimi K3 · high · 86.7% · US$ 0.48Kimi K3 · max · 94.5% · US$ 0.77
Claude Opus 4.8GPT-5.5Kimi K3

DeepSWE

The score scale does not start at zero.

20%35%50%66%81%US$ 2US$ 5US$ 10cost per task (log scale)Claude Opus 4.8 · low · 40.8% · US$ 2.29lowClaude Opus 4.8 · medium · 48.7% · US$ 3.44mediumClaude Opus 4.8 · high · 51.8% · US$ 4.28highClaude Opus 4.8 · xhigh · 54.4% · US$ 8.01xhighClaude Opus 4.8 · max · 59.0% · US$ 13.22maxGemini 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$ 2.18Gemini 3.8 Flash · medium · 71.0% · US$ 1.97Gemini 3.8 Flash · high · 73.8% · US$ 2.36GPT-5.5 · low · 27.0% · US$ 1.20GPT-5.5 · medium · 54.0% · US$ 2.75GPT-5.5 · high · 64.4% · US$ 5.10GPT-5.5 · xhigh · 67.0% · US$ 7.23
Claude Opus 4.8Gemini 3.7 FlashGemini 3.8 FlashGPT-5.5
Family

Other models from Anthropic

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

ModelReleaseECIELOLiveBenchAccess
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
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
Your rating

Have you used Claude Opus 4.8? 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 4.8

Does Claude Opus 4.8 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 May 28, 2027 on the platforms it operates, with at least 60 days' notice.

How much does Claude Opus 4.8 cost?

US$ 5 per million input tokens and US$ 25 per million output, the same as Opus 4.7. 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.

Why was a request to Fable or Opus 5.5 answered by Opus 4.8?

Because those models' safety classifiers blocked the request, usually a cybersecurity one, and the answer fell to the fallback model. In the Claude apps this happens by default, with a notice. On the API, only with the fallbacks parameter, and it is billed at Opus 4.8 rates.

Is it worth migrating from Opus 4.8 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 4.8 rank among AI models?

In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Opus 4.8 ranks #16 out of 267 models, with an ECI of 158.26. 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 4.8 better than GPT-5.5?

Across the 37 tests both models took and Epoch AI compiled, Claude Opus 4.8 comes out ahead on 9, behind on 28 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 4.8 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 4.8 are: Lmca (#4 of 65); FrontierMath tier 4 (#3 of 33); WeirdML (#5 of 57). 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 4.8 announcement
  2. Anthropic, Claude Opus 4.8 system card
  3. Anthropic, Claude Opus 4.8 overview (documentation)
  4. Anthropic, prompting Claude Opus 4.8 (documentation)
  5. Anthropic, pricing (documentation)
  6. Anthropic, model deprecations (documentation)
  7. Anthropic, migrating to Claude Opus 5.5 (documentation)
  8. Anthropic, refusals and fallback (documentation)
  9. Anthropic, platform release notes (documentation)
  10. Anthropic, Claude Opus 5 announcement (comparison with Opus 4.8)
  11. Anthropic, Claude Fable 5 announcement (fallback to Opus 4.8)
  12. Anthropic, Claude Opus 5.5 announcement (cybersecurity rerouting)

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