SeaWorkClaude Opus 5 for Coding (2026): 96% SWE-bench, $5/$25 Pricing & How to Run It
Claude Opus 5 release date, pricing, coding benchmarks, Fable 5 comparison, and how to control a Claude Code Opus 5 run from your phone with SeaWork.
Claude Opus 5 is Anthropic’s new coding model, announced on July 24, 2026. Its API model ID is claude-opus-5, it scored 96.0% on SWE-bench Verified, and its standard API price is $5 per million input tokens and $25 per million output tokens. Anthropic calls it a thoughtful, proactive model that comes close to Claude Fable 5’s frontier intelligence at half the price.
That matters for coding agents because a near-frontier model at half the price makes long agent runs affordable to leave running. A long run should not require staying at a desk: run Claude Code on the machine that holds your repository, then keep its work, tool calls, and approval requests within reach from your phone.
What Claude Opus 5 is
Claude Opus 5 is available now through Claude Code and the Claude API. Developers can start with the API model ID claude-opus-5.
Anthropic’s headline comparison is direct: Opus 5 comes close to the frontier intelligence of Claude Fable 5 at half the price. That is a useful framing, but it does not make Fable obsolete. Fable 5 is Anthropic’s model for ambitious, long-running, asynchronous work; use the benchmark and pricing differences below to decide what is worth paying for on a particular run. For the earlier release and its operating context, see Claude Fable 5 on your phone.
Claude Opus 5 benchmarks: what the numbers say
The system card shows a large step over Opus 4.8 across software-engineering and computer-use evaluations. Opus 5 is close to Fable 5 on several coding measures, leads it on some, and is behind it on others. The rows below use the published scores from Anthropic’s Opus 5 System Card.
| Benchmark | Opus 5 | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| SWE-bench Pro | 79.2 | 69.2 | 80.0 | 64.6 |
| SWE-bench Multilingual | 89.5 | 84.4 | 86.6 | - |
| SWE-bench Multimodal | 59.4 | 38.4 | 54.1 | - |
| DeepSWE v1.1 | 68.8 | 59.0 | 69.7 | 72.7 |
| FrontierCode 1.1 Main | 53.4 | 46.5 | 53.5 | 47.5 |
| FrontierBench v0.1 | 43.3 | 21.1 | 33.8 | - |
| OSWorld 2.0 | 70.6 | 55.7 | 66.1 | 62.6 |
| ARC-AGI-3 | 30.2 | 1.5 | - | 7.8 |
The 96.0% SWE-bench Verified result is stated in the system card’s prose, not in this summary table. For context, Anthropic reports 95% for Fable 5 and 95.5% for Mythos 5 on that benchmark.
Three comparisons are especially useful for coding decisions. Opus 5 scores 53.4 on FrontierCode 1.1 Main, just below Fable 5’s 53.5 and above GPT-5.6 Sol’s 47.5. On FrontierBench v0.1, it reaches 43.3, above Fable 5’s 33.8 and more than double Opus 4.8’s 21.1. And on CursorBench 3.2 at max effort, Anthropic says Opus 5 performs within 0.5% of Fable 5’s peak score at half the cost per task.
The widest gap in the card is ARC-AGI-3, a novel-problem-solving benchmark rather than a coding one: Opus 5 scores 30.2 at high effort against Opus 4.8’s 1.5 and GPT-5.6 Sol’s 7.8. Fable 5 has no entry. Treat that as a signal about unfamiliar reasoning problems, not about day-to-day repository work.
Benchmarks are evidence about standardized tasks, not a promise about your repository. Use them to narrow the choice, then judge the model on the codebase, test suite, tools, and review standards that actually matter to you. For a broader frontier-model comparison, read GPT-5.6 vs Claude Fable 5.
Pricing: half of Fable 5, same as Opus 4.8
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens at standard pricing, the same price as Opus 4.8. Claude Fable 5 and Claude Mythos 5 are priced at $10 input / $50 output per million tokens, so Opus 5 is literally half of Fable 5’s token price.
Anthropic also offers a fast mode for Opus 5: up to 2.5x faster speeds at 2x standard pricing. Whether that trade is worthwhile depends on whether your bottleneck is elapsed time or the cost of the run.
The practical implication is simple. If an Opus 5 run is capable enough for the task, its standard price leaves more room for long exploratory work than Fable 5. If you are comparing Claude access and billing paths more broadly, the Claude Code pricing guide lays out the surrounding decisions.
Opus 5 vs Fable 5 vs GPT-5.6: which one for coding
Choose Opus 5 when you want a high-end coding model at the Opus 4.8 price. Its published 96.0% SWE-bench Verified score, 43.3 FrontierBench v0.1 score, and near-Fable CursorBench result make it the sensible first model to try for a difficult implementation, debugging, or refactoring run.
Choose Fable 5 when its strengths are the right fit for the task and its higher price is justified. Anthropic describes Fable 5 as best for ambitious, long-running, asynchronous work. Fable released on June 9, 2026, was suspended under US export controls on June 12, and returned globally on July 1, 2026. Its API model ID is claude-fable-5.
Treat GPT-5.6 Sol as another benchmarked comparison point, not a substitute for testing. In the Opus 5 system-card table, Sol leads Opus 5 on DeepSWE v1.1, while Opus 5 leads Sol on SWE-bench Pro, FrontierCode 1.1 Main, and OSWorld 2.0. The right choice can differ by task. The useful workflow is to make the model selection explicit, keep tests as the gate, and move up a tier only when the task merits it.
Mythos 5 is not an everyday alternative: it is the same underlying model as Fable 5 with safeguards lifted, restricted to a small set of cybersecurity and biology testing partners. Anthropic describes Mythos-class models as a capability tier above Opus.
How to run Claude Opus 5 from your phone with SeaWork
Opus 5 is available in Claude Code today. SeaWork does not need to be the model gateway for this to be useful: it runs your Claude Code agent on your machine, then gives you a control surface from your phone and desktop. Set Claude Code to use Opus 5, and SeaWork lets you steer that run remotely. Your code, credentials, and MCP servers stay local.
The SeaWork gateway catalog does not currently list claude-opus-5, so do not expect SeaWork’s gateway model picker to provide it. This is about controlling the Claude Code session you already run, not routing Opus 5 through a SeaWork-hosted catalog. For the day-to-day mechanics of that agent workflow, see how to use Claude Code.
Start and follow a Claude-provider agent from the terminal:
seawork run --provider claude "refactor the payments module"
seawork ls -a
seawork logs <id> -fFrom the phone, you can watch every tool call live, read the full conversation, send a mid-run message to course-correct, and answer approval prompts when they arrive. That fits how a thoughtful, proactive model actually gets used: the longer a run goes unattended, the more it matters that you can see what it did and step in without returning to a desk.
When Opus 5 fits - and when to reach for Fable 5
Reach for Opus 5 first when you want the strongest published Opus 5 coding results at the lower price point. It is a clear candidate for a substantial task where you need capable autonomous work but want to keep cost below Fable 5’s $10/$50 rate.
Reach for Fable 5 when the task is especially ambitious or long-running and your own evaluation shows that its marginal capability is worth twice the token price. The system-card scores do not establish one permanent winner: Fable is slightly ahead on SWE-bench Pro, DeepSWE v1.1, and FrontierCode 1.1 Main, while Opus 5 is ahead on SWE-bench Multilingual, SWE-bench Multimodal, FrontierBench v0.1, and OSWorld 2.0.
For either model, the safe default is the same: begin with a scoped task, watch the tool trajectory, keep tests and review as the acceptance criteria, and intervene early when the agent chooses the wrong path.
FAQ
When was Claude Opus 5 released?
Anthropic announced Claude Opus 5 on July 24, 2026. Anthropic says it is available today, including for developers through the Claude API with the model ID claude-opus-5.
How much does Claude Opus 5 cost compared with Fable 5? Opus 5 costs $5 per million input tokens and $25 per million output tokens. Fable 5 costs $10 input and $50 output per million tokens, making Opus 5 half the price on both standard token rates. Opus 5 fast mode is up to 2.5x faster at 2x standard pricing.
What does Claude Opus 5’s 96.0% SWE-bench Verified score mean? It is Anthropic’s published score on SWE-bench Verified, reported in the prose of the Opus 5 System Card. It is a strong standardized software-engineering result, but it is not a guarantee that the model will solve a particular issue in your repository. Your own tests and review remain the deciding evidence.
What is the Claude Opus 5 context window? The system card evaluates context windows up to 1M tokens and says evaluation-dependent windows do not exceed 1M tokens. For work beyond that window, Anthropic used context compaction triggered at 200k tokens. Anthropic has not published a GA or beta label for the 1M-token context evaluation.
Do I need a separate API key to use Opus 5 with SeaWork?
SeaWork controls the Claude Code agent running on your machine; it does not make its gateway catalog serve claude-opus-5. Configure Opus 5 in the Claude Code environment you run, then use SeaWork to watch and steer that session. Your credentials remain local to that environment.
Claude Opus 5 vs Fable 5: which should I use for everyday coding? Start with Opus 5 when its published coding results match the task and you want half of Fable 5’s standard token price. Reach for Fable 5 when a task is ambitious or long-running enough that your own testing shows its extra capability is worth the higher rate.
Where to start
Claude Opus 5 is available now through Claude Code and the Claude API. Start with a bounded coding task, set the model to claude-opus-5, and use your tests to decide whether it is the right default for your work. When a run is long enough to leave the desk, SeaWork keeps its progress and approvals visible from your phone without moving the repository or credentials off your machine.
Download SeaWork to control the local Claude Code agent you already use, or compare the surrounding Claude workflow in the Claude Code pricing guide.