Claude Opus 5 review
Opus 5 is the best-value Claude for difficult, long-running agent work, but Sonnet 5 still wins on routine cost.
Anthropic released Claude Opus 5 on July 24, 2026 with a bold promise: intelligence close to Claude Fable 5 at half the API price. The useful question is not whether it tops a launch chart. It is whether the model finishes more real work per dollar than Opus 4.8 or Sonnet 5.
Our verdict is yes for difficult, long-running work. Opus 5 is the sensible upgrade from Opus 4.8 because the standard token price is unchanged, while public coding and professional-work evaluations show a large quality gain. It is less compelling for short edits, extraction, classification, or high-volume chat, where Sonnet 5 remains cheaper and usually fast enough.
Evidence note, checked July 26, 2026. This is an evidence-based review, not a claim that OmniaKey independently reran every benchmark. We compared Anthropic's release notes and technical documentation with the public Frontier-Bench leaderboard and Artificial Analysis' independent GDPval-AA v2 results. Vendor-run results are labeled as such.
Claude Opus 5 review: the short answer
| Your workload | Best starting point | Why |
|---|---|---|
| Multi-file features, difficult debugging, or long agent runs | Opus 5 at high effort | Best balance of capability and spend; raise to xhigh only when your evals justify it |
| Routine coding, support, extraction, or high-volume tasks | Sonnet 5 | Lower token price and faster comparative latency |
| A current Opus 4.8 integration | Test and move to Opus 5 | Same standard price, stronger public results; behavior changes still need regression tests |
| Maximum capability with cost secondary | Compare Opus 5 with Fable 5 | Fable has the higher advertised ceiling, but Opus costs half as much and is close or ahead on several public tasks |
The biggest improvement is not one-shot code generation. It is staying oriented through a long task: finding the root cause instead of patching the symptom, using tools until the result is verified, and completing a deliverable without leaving placeholders.
What changed from Opus 4.8
| Specification | Claude Opus 5 |
|---|---|
| API model ID | claude-opus-5 |
| Context window | 1 million tokens, both default and maximum |
| Maximum synchronous output | 128,000 tokens |
| Reliable knowledge cutoff | May 2026 |
| Standard API price | $5 / million input tokens; $25 / million output tokens |
| Prompt cache | $6.25 / million for a 5-minute write; $10 for a 1-hour write; $0.50 for a hit |
| Thinking | Adaptive thinking is on by default |
| Effort levels | low, medium, high, xhigh, and max; default is high |
The headline limits match Opus 4.8, but the operating behavior does not. Opus 5 turns thinking on by default, scales more reliably when effort increases, and lowers the minimum cacheable prompt from 1,024 to 512 tokens. It also supports beta mid-conversation tool changes, so an agent can add or remove tools without invalidating the existing prompt cache.
Fast mode is available as a research preview on the Claude API. Anthropic says it runs around 2.5 times the standard speed, but it also doubles the base price to $10 input and $50 output per million tokens. It is a latency purchase, not a discount.
Benchmarks: strong results with important footnotes
Frontier-Bench favors Opus 5 by a wide margin
The public Frontier-Bench v0.1 leaderboard lists Opus 5 with mini-SWE-agent at 43.53% +/- 1.65% resolution. The same table lists Fable 5 at 33.78% and Opus 4.8 at 21.08%. In other words, the posted Opus 5 result is about twice the Opus 4.8 result and roughly ten percentage points above Fable 5.
That is meaningful, but it is not a neutral laboratory result. Anthropic's footnote says this was an internal run on a GKE backend, averaged over five attempts per task. Opus 4.8 also served as the fallback when safety classifiers refused an Opus 5 or Fable 5 request. The benchmark supports the claim that Opus 5 is much better at hard agent work; it does not prove that every repository will see the same uplift.
GDPval-AA gives the strongest independent signal
Artificial Analysis' GDPval-AA v2 evaluates 220 real-work tasks across 44 occupations. Models operate in an agent loop with shell and web tools, then produce documents, spreadsheets, slides, diagrams, and other deliverables. Blind pairwise comparisons produce an Elo score.
| Effort and model | GDPval-AA v2 Elo | 95% interval shown by leaderboard |
|---|---|---|
Opus 5, max | 1861 | -25 / +25 |
Opus 5, xhigh | 1827 | -24 / +24 |
Fable 5, max | 1747 | -17 / +17 |
Opus 5, high | 1741 | -23 / +23 |
GPT-5.6 Sol, max | 1735 | -17 / +17 |
Opus 5, medium | 1632 | -22 / +22 |
Sonnet 5, max | 1603 | -17 / +17 |
Opus 4.8, max | 1593 | -16 / +16 |
Opus 5, low | 1454 | -20 / +20 |
Two conclusions matter. First, Opus 5 at max and xhigh leads this independent leaderboard by a clear margin. Second, effort is not cosmetic: the gap between low and max is 407 Elo points. At default high, Opus 5, Fable 5, and GPT-5.6 Sol have overlapping intervals, so a six-point ordering should not be sold as a decisive win.
Anthropic also reports that Opus 5 at max lands within 0.5% of Fable 5's peak CursorBench score at half the cost per task, triples the next-best score on ARC-AGI 3, and beats Fable 5's best OSWorld 2.0 result at just over one-third of the cost. Those are useful directional signals, but they come from the launch materials rather than an independent run, so we weigh them less heavily than GDPval-AA.
Where Opus 5 looks like a real upgrade
Long-horizon coding. Anthropic's prompting guide says the model is strongest on multi-file features, larger refactors, and end-to-end work. The Frontier-Bench result supports that positioning. The practical gain is fewer abandoned stubs and less human intervention midway through a task.
Debugging and code review. Early users report better root-cause analysis and fewer surface-only fixes. Anthropic also says review recall stays useful at lower effort. Teams should still measure false positives on their own code review set, because launch testimonials are not a substitute for local precision and recall.
Visual and document work. Opus 5 is positioned as stronger on charts, diagrams, frontend replication, spreadsheets, and slide decks. Its advantage should be largest when the agent can inspect the rendered result and iterate, not when it only emits code or markup once.
Self-verification. The model checks its work more often without being asked. That helps on open-ended tasks, but duplicated instructions such as "double-check everything" can make it over-verify, spend more tokens, and delegate too much.
The upgrade is smaller on a one-file edit with a precise acceptance test. A cheaper model can solve those tasks before Opus 5's deeper reasoning has any opportunity to pay for itself.
Claude Opus 5 pricing in a real task
For a task that sends 100,000 uncached input tokens and receives 10,000 output tokens, standard list-price arithmetic looks like this:
| Model | Standard input / output price | Example task cost |
|---|---|---|
| Claude Fable 5 | $10 / $50 per million | $1.50 |
| Claude Opus 5 | $5 / $25 per million | $0.75 |
| Claude Opus 4.8 | $5 / $25 per million | $0.75 |
| Claude Sonnet 5 | $3 / $15 per million | $0.45 list price |
Sonnet 5 has an introductory $2 / $10 rate through August 31, 2026, which makes that sample task $0.30 during the promotion. Prices above exclude tool fees and any gateway pricing.
The 1-million-token window is not free capacity. Filling it once costs $5 in input before the answer. Reading the same cached million tokens costs $0.50, so prompt caching changes the economics of repeated repository or knowledge-base work dramatically. A 5-minute cache write costs $6.25 per million tokens and a 1-hour write costs $10.
Token price also does not equal task price. Higher effort can add thinking tokens and tool calls, while a stronger model can finish in fewer turns. Benchmark your completed-task cost, not only the rate card. The current Claude Opus 5 model page shows OmniaKey's live gateway rate separately from Anthropic's standard price.
Who should upgrade
Opus 4.8 users: upgrade after a focused regression test. The standard price and context limit are unchanged, so there is little economic reason to stay on 4.8 when Opus 5 improves your own completion rate. Keep 4.8 pinned until you test output length, thinking behavior, tool calls, and latency.
Sonnet 5 users: do not move every request. Keep Sonnet for routine turns and escalate ambiguous bugs, architecture work, large refactors, or expensive failures to Opus 5. That routing policy usually captures more value than selecting one model for everything.
Fable 5 users: test Opus 5 first on cost-sensitive production work. Fable remains Anthropic's highest-capability generally available model, but the public evidence shows many workflows where Opus 5 is close, equal, or better at half the token price. Pay the Fable premium only where your evaluation set proves it matters.
Security teams: treat the general benchmark story carefully. Anthropic intentionally did not train Opus 5 as its offensive cyber model, and it remains behind Mythos 5 on exploit development. Finding a source-code vulnerability and developing a working exploit are different tasks.
Migration traps to test before switching
- Change the exact model ID from
claude-opus-4-8toclaude-opus-5. Dateless Claude 5 IDs are pinned snapshots, not silently moving aliases. - Revisit
max_tokens. Thinking is on by default, and the limit covers thinking plus visible output. - Do not combine
thinking: {"type": "disabled"}withxhighormax; the API returns HTTP 400. Keep thinking enabled and lower effort when cost is the concern. - Do not expect lower effort to shorten the visible answer reliably. Prompt for concise output separately.
- Remove inherited instructions that force extra verification or subagents. Opus 5 already does both more readily than Opus 4.8.
- Run an effort sweep on your own tasks. Anthropic recommends starting at the default
high, moving toxhighfor demanding agent work, and usinglowormediumwhere quality holds.
A sensible production rollout compares the same fixed task set at medium, high, and xhigh; records success, human correction time, latency, input, cache, output, and tool-call cost; then routes by task class. A single impressive transcript is not an evaluation.
Final verdict
Claude Opus 5 is the strongest default choice in Anthropic's lineup for developers whose expensive problems are long, ambiguous, and tool-heavy. It gives Opus 4.8 users a substantial published quality upgrade at the same standard price, and it makes Fable-level performance more affordable on several important workloads.
It is not the universal default. Sonnet 5 remains the better economic choice for routine work, max effort can erase savings through token use, and the launch-day evidence is still weighted toward Anthropic's own evaluations. Start at high, measure completed-task cost, and escalate selectively.
Frequently asked questions
Is Claude Opus 5 better than Opus 4.8?
Yes on the available public evidence. Frontier-Bench lists 43.53% resolution for Opus 5 versus 21.08% for Opus 4.8, and GDPval-AA lists 1861 Elo for Opus 5 at max versus 1593 for Opus 4.8 at max. The standard API token price is unchanged. Your prompts and agent harness still need a regression test.
Is Claude Opus 5 better than Sonnet 5?
Opus 5 is stronger for difficult reasoning and long agent work; Sonnet 5 is cheaper and faster for routine workloads. Use task-level routing rather than paying the Opus rate for every request.
How much does Claude Opus 5 cost?
Anthropic's standard API price is $5 per million input tokens and $25 per million output tokens. Cache hits are $0.50 per million tokens. Fast mode costs $10 input and $50 output per million.
Does Claude Opus 5 have a 1M context window?
Yes. The 1-million-token window is both the default and the maximum, with up to 128,000 synchronous output tokens. Client or gateway limits can still be lower.
What effort level should I use?
Start with high. Use medium when your evaluations show comparable quality at lower cost, xhigh for demanding coding or long-running agents, and max only when unconstrained token spending produces measurable value.