Release analysis
Opus 4.8: what the release benchmarks actually showed.
On 28 May 2026, Anthropic released Claude Opus 4.8 at the same regular API price as Opus 4.7: $5 per million input tokens and $25 per million output tokens. It supported a one-million-token context window. My question at launch was not whether every score rose. It was what the evaluation changes suggested about the kind of work Anthropic was improving.
Dated analysis. This essay records a release-day reading, updated on 13 July 2026 for corrected evaluation details. Anthropic later launched Fable 5 and Mythos 5 on 9 June and Sonnet 5 on 30 June. Those releases changed the available choices. What follows is not a current buying guide.
The deltas, not the scores
A benchmark is a controlled test, not a forecast for your workflow. A same-family comparison is still useful when the task, tools and scoring are held steady. Anthropic's system card reports the following figures. These are vendor-reported results, generally run with adaptive thinking at maximum effort and averaged across five trials unless noted.
| Anthropic-reported evaluation | 4.8 | 4.7 | Δ |
|---|---|---|---|
| SWE-bench Pro harder real issues | 69.2 | 64.3 | +4.9 |
| OSWorld-Verified computer use, revised harness | 83.4 | 82.8* | +0.6* |
| BrowseComp single-agent web search, rerun harness | 84.3 | 79.8 | +4.5 |
| MCP-Atlas tool use, updated configuration | 82.2 | 79.1 | +3.1 |
| SWE-bench Verified bug fixes | 88.6 | 87.6 | +1.0 |
| USAMO 2026 proof grading, 10 attempts per problem | 96.7 | 69.3 | +27.4 |
| GPQA Diamond science questions, 25 trials | 93.6 | 94.2 | −0.6 |
The corrected table weakens the neatest version of my original thesis. Agentic work did improve, especially on harder software issues and web search, but not every agent score jumped by five points. OSWorld moved only 0.6 points after Anthropic reran prior models with a zoom-tool fix and a larger per-turn token allowance. BrowseComp used a new blocklist, adaptive thinking and context compaction. MCP-Atlas used Scale AI's updated judge and a 100-tool-call budget. Those are cleaner within-table comparisons than the previously published numbers, but they also show how much the harness shapes a model score. Anthropic's revised system card reports 82.8 for the Opus 4.7 OSWorld rerun, while the current release-page footnote says 82.3. That unresolved official-source difference makes the delta either 0.6 or 1.1, so the table marks the system-card figure with an asterisk.
The knowledge story is mixed rather than flat. GPQA moved down 0.6 points, while USAMO rose sharply under a specialised setup: ten attempts per problem, high effort, a 300,000-token limit and model-based proof grading. That is evidence of better performance under those conditions, not 27.4 points of general intelligence. My inference is narrower: the release placed visible emphasis on sustained work with tools. The public results do not reveal how Anthropic allocated training effort, and they do not prove that knowledge stopped improving.
The honesty claim needs its conditions
Anthropic also reported that Opus 4.8 was around four times less likely than Opus 4.7 to let flaws in its own code pass without comment. The system card gives useful boundaries. One related test used short, prefilled coding transcripts that were not generated by the evaluated model, then checked whether a status summary disclosed failed tests, unfinished features or unsigned design decisions. Another used small code and data-analysis traps. Anthropic calls these tests simplified and less representative of long, real deployments.
This is a promising vendor result about disclosure, not proof that the model fixes its own mistakes or can be trusted without supervision. Independent evidence would require another party to reproduce the result, and local evidence requires your own workflow to prove the outcome with tests and review.
How I would evaluate the upgrade
Suppose a team uses an agent to update a dependency across several services. Build a set of twenty completed migrations from your own history. Give Opus 4.7 and 4.8 the same repository snapshot, instructions, tools, effort level and token budget. Run each task more than once. Score whether the existing tests pass, whether the agent adds a relevant regression test, whether it reports incomplete work, how much a reviewer must correct and what the run costs. Hide the model name from the reviewer where practical.
Decide from the pattern, not the launch table. If 4.8 reduces omitted failures on difficult migrations but costs more review time on routine changes, route only the difficult class to it. If the difference disappears under your harness, keep the cheaper or faster option. Product features such as Claude Code's dynamic workflows may use parallel agents, but that orchestration is separate from the base model and needs its own evaluation.
Evaluate a model upgrade inside your harness
Vendor benchmarks suggest where to look. A paired local test decides which work should move.
- Choose representative completed tasks Include ordinary migrations and difficult edge cases.
- Hold the harness steady Same repository state, prompt, tools, effort and token budget.
- Run repeated paired trials Reduce the influence of one lucky or unlucky run.
- Blind review where practical Score tests, regression coverage, omissions, correction time and cost.
- Route by observed pattern Use the newer model only for the task class where it earns the responsibility.
If the difference disappears under your conditions, keep the cheaper or faster option.
Better at a benchmark is not permission to leave
The system card also reports that Opus 4.8 was less robust than 4.7 in some agentic prompt-injection tests before Anthropic's safeguards closed the gap. A stronger agent can therefore increase both useful capability and the consequence of a bad instruction, compromised page or excessive permission. Keep consequential actions sandboxed, require deterministic tests, preserve rollback and place a person at the approval boundary.
The lasting lesson from Opus 4.8 is not “doing replaced knowing.” It is that model, tools, effort settings, context policy and evaluation harness form one system. Release benchmarks can suggest where to look. Your evidence decides whether the upgrade earns responsibility in your work.
Sources & caveats. Product facts and evaluation figures come from Anthropic's 28 May release and Opus 4.8 system card. Both are primary vendor sources, not independent replication. I excluded cross-vendor rankings whose harnesses were not configured alike. Later-release context comes from Anthropic's official Fable 5 and Mythos 5 and Sonnet 5 announcements.