On May 28, 2026, Anthropic shipped its next flagship, Opus 4.8 — gains in honesty, agentic coding, and million-token long context. Grounded in the official announcement and cloud model cards, with the prior version and rivals on the same table.
On May 28, 2026, Anthropic released its next flagship, Claude Opus 4.8. It lands just ~41 days after Opus 4.7, continuing a noticeably faster release cadence. The API model string is claude-opus-4-8; on AWS Bedrock it is anthropic.claude-opus-4-8.
This piece sticks to the facts — grounded in Anthropic's official materials and cloud model cards (primary sources) — covering what improved and how it benchmarks against the prior version and rivals. Some of Anthropic's official benchmark numbers are published as an image table rather than as body text; for those we cross-checked transcriptions from reputable secondary outlets, flagged under each table.
① "Honesty" is the headline. Anthropic positions Opus 4.8 as "more honest about its progress and able to work independently for longer." The most concrete figure is in coding: it is roughly 4× less likely to pass off flaws in its own code without flagging them, and it makes fewer unsupported claims while surfacing uncertainty.
② Alignment and safety metrics improved. Anthropic's alignment team reports that misaligned behavior (deception, helping with misuse) dropped substantially vs 4.7. A reported quantitative figure: a misalignment score of ~2.5 (4.7) down to ~1.9 (4.8) across thousands of simulated sessions per model (secondary sources).
③ Big gains in agentic coding and long context. As the tables show, the largest jumps are in SWE-bench Pro and 1M-token long-context tracking (GraphWalks). Default effort is now set to high, so it does more with a similar token budget to 4.7.
④ Speed and cost. The new Fast mode is ~2.5× faster than standard and ~3× cheaper than the previous generation's fast mode. Standard pricing matches 4.7 ($5 input / $25 output per 1M tokens).
These are Anthropic's official benchmark figures (image table, cross-checked via secondary transcriptions). Same methodology and same vendor, so this is the most reliable comparison.
| Benchmark | Opus 4.8 | Opus 4.7 | Δ |
|---|---|---|---|
| SWE-bench Verified (coding) | 88.6 | 87.6 | +1.0 |
| SWE-bench Pro | 69.2 | 64.3 | +4.9 |
| SWE-bench Multilingual | 84.4 | 80.5 | +3.9 |
| Terminal-Bench 2.1 (terminal) | 74.6 | 66.1* | +8.5 |
| OSWorld-Verified (computer use) | 83.4 | ~82.3 | +1.1 |
| GraphWalks BFS (1M context) | 68.1 | 40.3 | +27.8 |
| GraphWalks Parents (1M context) | 83.3 | 56.6 | +26.7 |
| USAMO 2026 (math) | 96.7 | 69.3 | +27.4 |
| HLE (no tools) | 49.8 | 46.9 | +2.9 |
| HLE (with tools) | 57.9 | 54.7 | +3.2 |
| GPQA Diamond (grad reasoning) | 93.6 | 94.2 | -0.6 |
| GDPval-AA (real-work ELO) | 1890 | 1753 | +137 |
| MCP-Atlas | 82.2 | 79.1 | +3.1 |
| Finance Agent v2 | 53.9 | 51.5 | +2.4 |
Units: % (except GDPval-AA, an ELO score). *Terminal-Bench versions differ (4.8 = 2.1 vs 4.7 = 2.0), so treat that delta with care. GPQA dipped slightly — a saturated benchmark in the low-to-mid 90s, so many read this as measurement noise. The jumps in USAMO and GraphWalks best capture the math and long-context gains.
As of May 2026, the competing frontier is OpenAI GPT-5.5 (released Apr 2026) and Google Gemini 3.1 Pro (released Feb 2026). The table below transcribes the comparison values published alongside Opus 4.8's official table. Caveat: each vendor uses different harnesses and conditions, so same-named benchmarks are not strictly 1:1 comparable. N/A = unpublished or not comparable.
| Benchmark | Opus 4.8 | Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-bench Pro | 69.2 | 64.3 | 58.6 | 54.2 |
| SWE-bench Verified | 88.6 | 87.6 | N/A | 80.6 |
| Terminal-Bench 2.1 | 74.6 | 66.1 | 78.2† | 70.3 |
| GPQA Diamond | 93.6 | 94.2 | N/A | 94.3 |
| OSWorld-Verified | 83.4 | ~82.3 | 78.7 | 76.2 |
| GraphWalks BFS (1M) | 68.1 | 40.3 | 45.4 | N/A |
| HLE (no tools) | 49.8 | 46.9 | 41.4 | 44.4 |
| HLE (with tools) | 57.9 | 54.7 | 52.2 | 51.4 |
| MCP-Atlas | 82.2 | 79.1 | 75.3 | 78.2 |
| Finance Agent v2 | 53.9 | 51.5 | 51.8 | 43.0 |
| GDPval-AA (ELO) | 1890 | 1753 | 1769 | 1314 |
Bold = best in row. †GPT-5.5's Terminal-Bench value ranges 78.2–83.4 across sources (harness-dependent), so it is hard to pin down. GPT-5.5's SWE-bench "Verified" is marked N/A here due to variant discrepancies. On the independent composite Artificial Analysis Intelligence Index v4.0, Opus 4.8 ranks #1 at 61.4 (GPT-5.5 60.2, Opus 4.7 57.3) — a third-party check.
Opus 4.8 leads its comparison set on agentic coding, computer use, 1M long-context tracking, and real-work agents. On pure knowledge reasoning (GPQA) it is effectively tied with Gemini 3.1 Pro; on terminal agents it is neck-and-neck with GPT-5.5. — Benchmarks, summed up
The Opus line has gone 4.5 (Nov 2025) → 4.6 (Feb 2026) → 4.7 (Apr 2026) → 4.8 (May 2026), with shrinking gaps. Reports the same week said Anthropic raised a large round and teased a higher-tier model, "Claude Mythos," as imminent. But those points (raise size, Mythos timing) are press reports and teasers, not confirmed by Anthropic's primary documents, so we don't assert them here.