AnthropicPrevious version · Claude Opus 5

Claude Opus 4.8

Anthropic's own docs already file Claude Opus 4.8 under "Legacy models" with a migration guide pointing to Opus 5 — but the demotion is mostly cosmetic: it's still formally Active until at least 2027-05-28, priced identically to its successor, and remains the model Anthropic's own API docs use as the worked fallback example when Claude Fable 5's safety classifiers refuse a request.

Claude Opus 4.8 benchmarks and pricing, every number sourced: 15 tracked Claude Opus 4.8 benchmark scores (14 independently run, 1 still resting on a vendor’s own claim), priced at $5.00 per million input tokens and $25.00 per million output.

Released
2026-05-28
License
proprietary
Context window
1M tokens
Knowledge cutoff
Not disclosed
Verified
sources checked 2026-08-13–2026-10-08
Parameters
Not disclosed
Architecture
Not disclosed

Claude Opus 4.8’s verified record

Claude Opus 4.8’s featured comparison is Gemini 3.1 Pro Preview: 2 leads, 2 ties, and 7 not callable. Claude Opus 4.8 is priced at $5.00/$25.00 per 1M tokens (in/out) vs Gemini 3.1 Pro Preview’s $2.00/$12.00. Full verdict →

Against the 320 head-to-head comparisons Claude Opus 4.8 shares with other tracked models:

77 real gaps58 inside the band185 not called

A gap counts for Claude Opus 4.8 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Claude Opus 4.8 trails on 43 of them.

HLE · no tools Reasoning

±2 is noise
Behind
Kimi K3.1 Pro −3.8 · GPT-6.1 Sol −4.2 · GPT-6 Astra −6.0 · Claude Sonnet 5.5 −6.3 · Gemini 4 Argon −8.4 · Claude Opus 5.5 −12.7 · 3 previous: Claude Opus 5 −6.2 · Claude Fable 5 −6.8 · Claude Fable 5.1 −10.4
Ahead
Step 5 Preview +2.2 · Claude Haiku 5.5 +4.3 · Muse Spark Open +4.8 · DeepSeek V4.1 Flash +5.4 · Qwen3.8-Max +5.6 · Grok 4.7 +5.6 · GLM-5.3 +6.4 · DeepSeek V4 Pro (0813) +7.7 · GLM-5.3-Flash +8.8 · MiniMax M3 +9.7 · GPT-6 Luna +10.2 · Qwen3.8-Flash-Next +10.7 · Mistral Large 4 +13.7 · Beam 1.0 +14.0 · Nemotron 3.5 Super VL +18.0 · 6 previous: Muse Spark 1.2 +3.2 · Grok 4.6 +5.8 · Claude Sonnet 5 +7.4 · GLM-5.2 +7.6 · GPT-5.6 Luna +9.2 · DeepSeek V4 Flash (0731) +10.1
Tie
10 models within ±2
Unverified
1 model — vendor-reported on one side

LiveBench Composite score across 7 domains

±2.7 is noise
Behind
Kimi K3 −3.0 · DeepSeek V4.1 Flash −4.9 · GPT-6.1 Sol −5.4 · Claude Opus 5.5 −7.0 · 5 previous: GPT-6 Sol −3.1 · Claude Opus 5 −3.9 · GPT-5.6 Sol −4.8 · Claude Fable 5 −6.8 · Claude Fable 5.1 −7.2
Ahead
GPT-6 Luna +4.2 · GLM-5.3-Flash +4.6 · MiniMax M3 +8.9 · 1 previous: GLM-5.2 +3.0
Tie
7 models within ±2.7
Setup-dependent
9 models — scored on a different harness or effort tier

Terminal-Bench 4.0 · max Terminal ops

±12.4 is noise
Behind
GLM-5.3 −15.7 · GPT-6.1 Sol −31.8 · GPT-6 Astra −36.4 · Claude Sonnet 5.5 −40.9 · Claude Opus 5.5 −41.9 · 5 previous: GPT-5.6 Sol −14.7 · Claude Fable 5 −18.2 · GPT-6 Sol −21.2 · Claude Opus 5 −30.3 · Claude Fable 5.1 −34.8
Ahead
1 previous: Claude Sonnet 5 +13.6
Tie
6 models within ±12.4
Unverified
1 model — vendor-reported on one side
Setup-dependent
1 model — scored on a different harness or effort tier

DeepSWE Long-horizon coding

±9.5 is noise
Behind
Kimi K3 −10.0 · GLM-5.3 −10.0 · Gemini 3.8 Flash −15.0 · GPT-6 Astra −15.0 · 3 previous: Claude Fable 5 −11.0 · GPT-5.6 Sol −14.0 · Claude Opus 5 −15.0
Ahead
1 previous: GLM-5.2 +15.0
Tie
8 models within ±9.5
Unverified
19 models — vendor-reported on one side
Setup-dependent
2 models — scored on a different harness or effort tier

AnalystAgent Spreadsheet & document analysis

±11.2 is noise
Behind
Claude Opus 5.5 −11.3 · Claude Sonnet 5.5 −12.5 · 1 previous: Claude Fable 5.1 −12.5
Ahead
GLM-5.3-Flash +20.0 · MiniMax M3 +35.0
Tie
17 models within ±11.2
Setup-dependent
5 models — scored on a different harness or effort tier

Agents' Last Exam Professional work

±3.2 is noise
Behind
GPT-6 Astra −7.2 · Claude Opus 5.5 −11.2
Ahead
1 previous: GLM-5.2 +6.6
Tie
2 models within ±3.2
Unverified
7 models — vendor-reported on one side
Setup-dependent
9 models — scored on a different harness or effort tier

ARC-AGI-2 · high Compositional visual reasoning

±9.2 is noise
Behind
Gemini 3.8 Flash −17.1 · Claude Opus 5.5 −21.2 · 1 previous: Gemini 3.7 Flash −12.5

Toolathlon-Verified Multi-tool chores

±9.7 is noise
Ahead
Gemini 3.1 Pro Preview +15.1 · 1 previous: GLM-5.2 +16.3
Tie
6 models within ±9.7
Unverified
11 models — vendor-reported on one side

Terminal-Bench 2.1 Terminal ops

±10.6 is noise
Tie
2 models within ±10.6
Unverified
3 models — vendor-reported on one side
Setup-dependent
29 models — scored on a different harness or effort tier

OSWorld 2.0 · v2026 06 24 full Long-horizon computer use

±9.7 is noise
Ahead
Unverified
1 model — vendor-reported on one side

No verdict for Claude Opus 4.8 anywhere on GPQA Diamond, LiveCodeBench, SWE-bench Verified ( saturated); HLE · with tools ( nothing independently confirmed on both sides); HMMT ( contaminated).

Claude Opus 4.8 API pricing

$5.00 in / $25.00 out per 1M tokens — official pricing source

What Claude Opus 4.8 costs per job

Claude Opus 4.8 cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.750
A codebase review1,000K / 100K$7.50
A day of agent work10,000K / 1,000K$75.00

Computed from Claude Opus 4.8’s list rates above — cache discounts and batch tiers are not applied.

Claude Opus 4.8 is one of 9 Anthropic models tracked on this site, at these official list prices.

Anthropic model pricing, official list rates
ModelIn / 1MOut / 1M
Claude Fable 5.2$8.00$40.00
Claude Haiku 5.5$0.10$0.50
Claude Sonnet 5.5$2.00$10.00
Claude Opus 5.5$4.00$20.00
Claude Fable 5.1(previous version)$10.00$50.00
Claude Opus 5(previous version)$5.00$25.00
Claude Sonnet 5(previous version)$2.00$10.00
Claude Fable 5(previous version)$10.00$50.00
Claude Opus 4.8(previous version)$5.00$25.00

Claude Opus 4.8 benchmark scores

Claude Opus 4.8 benchmark scores, provenance, and source links
BenchmarkScore
HLE(no tools)[1]
Reasoning · ±2 is noise
HLE(with tools)[2]
Reasoning · ±2 is noise
Terminal-Bench 2.1[3]
Terminal ops · ±10.6 is noise
DeepSWE[4]
Long-horizon coding · ±9.5 is noise
Toolathlon-Verified[5]
Multi-tool chores · ±9.7 is noise
Agents' Last Exam[6]
Professional work · ±3.2 is noise
SWE-bench Verifiedsaturated[7]
Bug fixing — not ranked at any gap size
GPQA Diamondsaturated[8]
Expert science Q&A — not ranked at any gap size
LiveCodeBenchsaturated[9]
Contest coding — not ranked at any gap size
ARC-AGI-2(high)[10]
Compositional visual reasoning · ±9.2 is noise
LiveBench[11]
Composite score across 7 domains · ±2.7 is noise
HMMTcontaminated[12]
Competition mathematics — not ranked at any gap size
AnalystAgent[13]
Spreadsheet & document analysis · ±11.2 is noise
Terminal-Bench 4.0(max)[14]
Terminal ops · ±12.4 is noise
OSWorld 2.0(v2026 06_24_full)[15]
Long-horizon computer use · ±9.7 is noise

Who ran these numbers: 14 of 15 independent — vals.ai (4), artificialanalysis.ai (2), tbench.ai (1), deepswe.datacurve.ai (1), toolathlon.xyz (1), snorkel.ai (1), arcprize.org (1), livebench.ai (1), matharena.ai (1), osworld-v2.xlang.ai (1); vendor self-reported (1).

  1. HLE: AA's own run, 'Claude Opus 4.8 (max)' — text-only, no tools; the board's underlying value is 48.66. Replaces the self-reported 49.8 from DeepSeek's comparison chart. Distinct from 'Fable 5 (Opus 4.8 fallback)' at 55.5 — a different system. Corrected 2026-10-03: this row previously cited the evaluations URL with ?models=claude-opus-4-8, but that query parameter is client-side only — the server returns the default payload byte-identical to the unfiltered URL, which no longer lists the superseded model, so the citation pointed at a page that does not carry the number. The value itself was re-confirmed against AA's model page (raw 0.486562), which is what is cited now.
  2. HLE: As published in DeepSeek's comparison chart, not independently verified against Anthropic's own materials.
  3. Terminal-Bench 2.1: Re-verified 2026-09-03 in a browser after npm run check-sources flagged this row: tbench.ai shipped Terminal-Bench 4.0 as the new default board around 2026-08-27 and rebuilt the site so version is a client-side selection — the old deep link now 308s to a URL whose server-rendered HTML always carries the 4.0 board, and the 2.1 board only loads after picking "2.1" from the on-page selector, which is why the static check reads MISS. Value confirmed unchanged: 78.9%, rank 5 of 17, 'Opus 4.8 (high)' via Claude Code (CI now shown as ±2.6%, not ±1.3 — recomputed in the redesign, not a re-run). Vendor chart claimed 85.0. AA's own different-harness run gives 84.6 — harness choice alone spans 6 points here. Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists two 'Claude Opus 4.8' rows, 71.91 (priced $5/$25) and 69.66 (no price shown).
  4. DeepSWE: 59%±2 pass@1, mini-swe-agent harness, rank 10 (board updated 2026-08-13). Replaces the self-reported 58.0 — the two agree within error.
  5. Toolathlon-Verified: Label correction, same value: the board's 76.2±3.4 (rank 2) carries Toolathlon's 'Evaluated by us' badge — this was always an independent run, we had it mislabeled self-reported.
  6. Agents' Last Exam: Overall pass rate 27.0 (Claude Code, Max; score 45.1; $3,985). Replaces 25.7 — that value was the ALE-CLI split (and coincidentally the launch chart's claim); normalized to Overall + independent. Per-effort results on 2026-10-01: Max 27.0, XHigh 22.4, High 20.4.
  7. SWE-bench Verified: vals.ai run, rank 9/83 (updated 2026-08-14). The board lists a second cheaper-config Opus 4.8 row at 85.8; 88.6 is the primary entry.
  8. GPQA Diamond: vals.ai run, rank 14/133, displayed tied with DeepSeek V4 Pro 0813 (updated 2026-08-15).
  9. LiveCodeBench: vals.ai run, rank 9/138 (updated 2026-08-15).
  10. ARC-AGI-2: Claude Opus 4.8's official ARC-AGI-2 leaderboard row, dated 2026-06-01 on arcprize.org. The model's Max-tier cell is genuinely blank (N/A) for ARC-AGI-2, so High is its usable ceiling.
  11. LiveBench: Board row "Claude 4.8 Opus Thinking Max Effort" on the 2026-06-25 LiveBench release.
  12. HMMT: MathArena's leaderboard row is "Claude-Opus-4.8 (max)", 95.45% across all 33 HMMT Feb 2026 problems (live-verified 2026-08-25 at matharena.ai/?comp=hmmt--hmmt_feb_2026). Carries MathArena's own contamination-warning flag — "Model was released after competition release" — a disclosed risk that applies to every one of this site's 4 tracked HMMT Feb 2026 rows, Claude Opus 4.8 included, since all 4 tracked models were released after HMMT Feb 2026 took place.
  13. AnalystAgent: AA's own run, board row 'Claude Opus 4.8 (Adaptive Reasoning, Max Effort)'. pass@1 63.5, pass@5 78.75 — the widest ceiling-versus-reliability spread among the Claude rows.
  14. Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 23.232 at effort max. Official board (grant-funded, Claude Code, max effort, 5-trial pass@1): 23.64 ± 3.56 CI — 0.4 points above this run.
  15. OSWorld 2.0: Authors' own run on the paper's original task set (the results file leaves the release column empty for this batch; values match the paper's headline config). Partial reward 54.8, binary accuracy 20.6. Maximum thinking, batched tool calls, 500 steps. The same release's standard-tool-call configuration scored 49.33 partial / 18.52 binary. Verified 2026-10-08 via direct fetch of the results JSON and a second read through a reader proxy — values identical.

Notes on the record

A previous version since Claude Opus 5 shipped (2026-07-24, same $5/$25 pricing). Anthropic docs move it to 'Legacy models' with a migration guide, but it is formally still Active — tentative retirement not before 2027-05-28 — and serves as Claude Fable 5's automatic fallback for classifier-refused requests. Standard pricing shown; a $10/$50 'Fast Mode' tier is also offered. Confirmed current as of 2026-08-20: Anthropic's official Opus 5 announcement (https://www.anthropic.com/news/claude-opus-5) verifies the identical $5/$25 standard rate and the 2026-07-24 launch date.

Anthropic's refusals-and-fallback API docs (https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback) confirm Claude Opus 4.8's fallback role with a worked example: a response object shows a Claude Fable 5 refusal handed off with "from": claude-fable-5, "to": claude-opus-4-8. That same page states the category-to-fallback-model mapping used by default routing is not published, so which specific refusal categories route to Opus 4.8 cannot be confirmed from this source.

Fast Mode runs at up to 2.5x speed; Anthropic's 2026-05-28 launch post says this is three times cheaper than previous fast-mode pricing but does not name the prior price or generation — third-party pricing trackers cite roughly $30/$150 on the Opus 4.7 generation, but that figure is unconfirmed in Anthropic's own materials and should be verified before being treated as fact.

Compare with

FAQ

Has Claude Opus 4.8 been independently benchmarked?

Mostly yes. 14 of the 15 benchmark results tracked for Claude Opus 4.8 are independent third-party runs: Artificial Analysis (HLE without tools, AA-AnalystAgent), tbench.ai (Terminal-Bench 2.1), DataCurve (DeepSWE), toolathlon.xyz (Toolathlon-Verified), Snorkel (Agents' Last Exam), Vals AI (SWE-bench Verified, GPQA Diamond, LiveCodeBench, Terminal-Bench 4.0), ARC Prize (ARC-AGI-2), LiveBench's own board, MathArena (HMMT Feb 2026), and the OSWorld 2.0 authors' own run. The one exception is HLE with tools (57.9, self-reported, observed 2026-08-13) — the 57.9 comes from DeepSeek's own launch comparison chart, neither Anthropic's nor an independent run's — a competitor's claim about a rival's score, which is why it can never earn a Real-gap or Tie label here.

Is Claude Opus 4.8 still available now that Claude Opus 5 has launched?

Yes. Anthropic's docs have moved Claude Opus 4.8 into a 'Legacy models' section with a migration guide toward Opus 5, but its formal lifecycle status is still Active, with a tentative retirement no earlier than 2027-05-28. It also has a live operational role beyond the docs label: Anthropic's refusals-and-fallback API documentation uses Claude Opus 4.8 as the worked example for Claude Fable 5's classifier-refusal fallback routing, including a response example that hands a refused request off from claude-fable-5 to claude-opus-4-8.

How does Claude Opus 4.8 perform against Claude Opus 5 on the benchmarks this site tracks?

Opus 5 leads on four of the eleven shared benchmarks, ties on one; six don't resolve. On DeepSWE (74.0 vs 59.0, +15.0), Terminal-Bench 4.0 (53.54 vs 23.23, +30.3, vals), Humanity's Last Exam without tools (54.9 vs 48.7, +6.2) and LiveBench (80.1 vs 76.2, +3.9), Opus 5 has four confirmed real gaps. Agents' Last Exam (32.2 vs 27.0) was real until 2026-10-02, but it pairs Opus 5's High run against Opus 4.8's Max, so it is setup-dependent now; on AA-AnalystAgent (53.75 vs 45.0 pass^5) the two are a statistical tie, inside the benchmark's noise band. Three more are graded saturated here — SWE-bench Verified, GPQA Diamond and, since 2026-09-29, LiveCodeBench (89.03 vs 87.82) — where both score near the ceiling, so none of those gaps means anything regardless of its size. The ninth, Terminal-Bench 2.1, is independently run on both sides but through different evaluation harnesses (tbench.ai's official board for Opus 4.8, Artificial Analysis for Opus 5), which this site treats as unfair. ARC-AGI-2, the tenth shared benchmark, is the same tier problem: 72.1 at High against 90.4 at Max, no matched pair. At an identical $5/$25 list price, the comparisons that do resolve favor Opus 5 — the case for staying on Opus 4.8 is inertia (an existing integration already pointed at it), not price or measured capability. Opus 4.8 isn't going away regardless: Anthropic's own infrastructure still relies on it as Claude Fable 5's automatic fallback.

How does Claude Opus 4.8's pricing compare with Claude Opus 5's?

They're identical at standard rates: $5 per million input tokens and $25 per million output tokens for both. Opus 5 launched 2026-07-24 at parity with Opus 4.8 rather than at Fable 5's higher $10/$50 tier, per Anthropic's official Opus 5 announcement, which explicitly states its pricing is 'the same as Opus 4.8.' That means moving from Claude Opus 4.8 to Opus 5 costs nothing extra at list price — the case for switching rests on capability and Anthropic's own default routing, not budget.

What does Claude Opus 4.8's Fast Mode cost and what does it change?

Fast Mode runs Claude Opus 4.8 at up to 2.5x the base speed for $10 per million input tokens and $50 per million output tokens — a fixed premium over the $5/$25 standard rate. Anthropic's own launch announcement describes this as three times cheaper than previous fast-mode pricing, though it doesn't state the exact prior price or name which earlier generation it's compared against. Third-party pricing trackers cite roughly $30/$150 on the Opus 4.7 generation, but that specific figure isn't confirmed in Anthropic's own materials.

Why does Claude Opus 4.8 handle Claude Fable 5's refused requests?

When Claude Fable 5's built-in safety classifiers decline a request, Anthropic's API supports routing that refusal to another model automatically — either to whichever model Anthropic recommends for the refusal's category (server-side default routing) or to a fallback list you configure yourself. Anthropic's refusals-and-fallback API reference uses Claude Opus 4.8 as its worked example for both paths, including a response example that shows the handoff as 'from': claude-fable-5, 'to': claude-opus-4-8. The docs state the exact category-to-model mapping used by default routing isn't published, so this confirms Opus 4.8's documented fallback role without pinning down which specific refusal categories send traffic to it. That's a production dependency that exists independently of Opus 4.8's 'Legacy' label in the docs navigation.

Further reading

Benchmark guides