Claude Opus 5 gets close to Fable 5 at half the price
Claude Opus 5 aims to approach Fable 5 capability at half the price. We examine its benchmarks, 1M context, effort control, availability and upgrade from Opus 4.8.
Anthropic has released Claude Opus 5, a model built for difficult agentic coding, document-heavy work and multi-step enterprise tasks. The important change is not the version number. Anthropic is trying to bring capabilities associated with the expensive Claude Fable 5 into a model that can be used every day without the same bill.
According to the official Anthropic announcement, Opus 5 approaches Fable 5 on many tasks while costing half as much. API pricing is $5 per million input tokens and $25 per million output tokens, exactly the same price as the previous Opus 4.8.
The model launched on July 24, 2026. It is available in Claude, Claude Code, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. It is now the default model for Claude Max subscribers and the strongest option included with Claude Pro.
Claude Opus 5 at a glance
| Feature | Claude Opus 5 |
|---|---|
| API model ID | claude-opus-5 |
| Launch date | July 24, 2026 |
| API price | $5 / 1M input tokens, $25 / 1M output tokens |
| Context window | 1 million tokens |
| Maximum output | 128k tokens |
| Training data and reliable knowledge cutoff | May 2026 |
| Main use cases | agentic coding, analysis, knowledge work, automation and research |
| Availability | Claude, Claude Code, API, AWS, Google Cloud and Microsoft Foundry |
| Fast mode | roughly 2.5 times faster at twice the base price |
The specifications are substantial, but context length alone does not separate Opus 5 from Anthropic's other recent models. Claude Fable 5 and Claude Sonnet 5 also support one million tokens. The difference is meant to appear in how the model plans, tracks progress and scales test-time compute to the difficulty of a task.
Near-Fable capability for everyday work
The Claude lineup increasingly resembles a set of tools for distinct classes of work. Fable 5 remains the option for the hardest assignments that an agent may run autonomously for hours or days. Sonnet 5 targets speed, scale and price-sensitive production. Opus 5 sits between them: capable enough for complicated work, but efficient enough to avoid becoming an emergency-only model.
| Model | Input / output price per 1M tokens | Context | Anthropic's intended role |
|---|---|---|---|
| Claude Fable 5 | $10 / $50 | 1M | hardest, long-running autonomous tasks |
| Claude Opus 5 | $5 / $25 | 1M | complex agentic coding and enterprise work |
| Claude Sonnet 5 | $3 / $15* | 1M | speed, scale and everyday integrations |
| Claude Opus 4.8 | $5 / $25 | 1M | previous Opus generation |
* Anthropic is offering lower introductory Sonnet 5 pricing through August 31, 2026.
This segmentation makes practical sense. The most capable model is not automatically the best choice for every request. If a routine task consumes several thousand tokens, the quality difference between Fable and Opus may not justify twice the price. In an agent workflow that processes millions of tokens, the cost gap compounds quickly.
Opus 5 benchmarks: strong claims with an important caveat
Anthropic presents Opus 5 as a new benchmark leader in coding and knowledge work. Its launch material says:
- on Frontier-Bench v0.1, the model ranks first and more than doubles Opus 4.8's performance at a lower cost per task;
- on CursorBench 3.2, maximum effort puts it within 0.5 percentage points of Fable 5's peak score at half the cost per task;
- on ARC-AGI 3, it scores three times higher than the next-best model;
- on Zapier's AutomationBench, its pass rate is about 1.5 times that of the next-best model at a comparable task cost;
- on OSWorld 2.0, it exceeds Fable 5's best result at a little over one-third of the cost.
Those results sound decisive, but methodology matters. Most of the comparison comes from vendor launch material, and performance changes with the effort setting, agent harness, number of attempts and cost calculation. On Frontier-Bench, safety-classifier refusals from Opus 5 could fall back to Opus 4.8. That does not make the evaluation useless, but it means the measured object was a deployed system rather than only the raw Opus 5 model.
The defensible conclusion is narrower: Opus 5 appears to be a major upgrade over Opus 4.8, but teams should confirm the advantage on their own repositories, documents and workflows.
Effort control changes the cost equation
One of the most useful additions is greater control over how much work the model performs. The effort parameter ranges from low to max. Lower settings reduce reasoning-token use and response time, while higher settings give the model more room to plan, verify and revise.
This is not merely a speed switch. A well-designed agent can use a different level at each stage:
lowfor request classification or file discovery;mediumfor straightforward edits and initial planning;highfor implementation, analysis and testing;xhighormaxfor difficult debugging and high-impact decisions.
Teams do not have to pay for maximum reasoning on every step. At the same time, effort should not be treated as a quality slider that always belongs at the top. Longer reasoning costs more and can add latency without improving routine work.
What Opus 5 changes in Claude Code
Opus 5 is positioned as particularly strong on long-running software tasks: understanding large repositories, finding root causes, editing multiple modules and completing several rounds of verification. Anthropic says the model is better at checking its own output and changing course when its first approach fails.
That distinction matters. Coding agents rarely fail because they have forgotten syntax. They fail because they fix a symptom instead of the cause, miss a dependency or declare success before running the relevant tests. Opus 5 is designed to improve the stretch between writing code and deciding that the job is actually finished.
None of this removes the need for ordinary safeguards. Terminal access, secrets, production databases and deployment permissions should remain constrained. Sandboxing, review, automated tests and execution budgets matter more than a leaderboard result.
Our evergreen Claude Code guide covers the tool, permissions and a sensible working environment in more detail.
A one-million-token context is not perfect memory
Opus 5 accepts up to one million input tokens and can produce as many as 128,000 output tokens in a synchronous API request. That capacity can hold a large repository, documentation, reports and a substantial history of previous decisions.
Capacity is not the same as attention. A model may technically accept a huge collection of material and still need clear goals, priorities and success criteria. Without structure, a million tokens becomes an expensive storage area where the most important fact is buried among files.
Long-context tasks work better with:
- a concise objective and constraints at the beginning;
- an index of the most important files or documents;
- an explicit definition of the finished result;
- periodic progress summaries;
- tests or examples that make the answer verifiable.
Safety: fewer interventions than Fable 5
Anthropic describes Opus 5 as its most aligned model to date. In the company's automated behavioral audit, it recorded the lowest rate of misaligned behavior among its recent systems.
Opus 5 is not intended to replace Mythos 5 for the most advanced biology and cybersecurity work. It can find vulnerabilities in source code, but stronger restrictions apply to penetration testing, binary scanning and exploit generation. In Claude and Claude Code, flagged requests may automatically fall back to Opus 4.8. Developers can enable similar fallback behavior through the API.
Anthropic expects Opus 5's classifiers to intervene about 85% less often than the safeguards applied to Fable 5. That matters for legitimate defensive work because overly broad refusals can block analysis during a real incident. Teams should still test the exact workflow rather than assume every security-related request will be accepted.
Should Opus 4.8 users migrate?
For new projects, the answer will often be yes. The base price has not changed, while Opus 5 is meant to improve coding, analysis, tool use and token efficiency. The claude-opus-5 identifier is a pinned model snapshot rather than an automatically moving alias.
Existing systems still need regression testing. Thinking is enabled by default in Opus 5, response verbosity can differ, and some tool-use and effort behavior has changed from Opus 4.8. A sensible evaluation should compare representative tasks and measure not only accuracy, but also token consumption, latency, tool calls and the share of outputs that require human correction.
The practical takeaway
Claude Opus 5 is not simply "the strongest Claude." Fable 5 still occupies the top tier of Anthropic's lineup. Opus 5 is more interesting as an attempt to move much of that quality into everyday coding, analysis and automation without doubling the bill.
If the vendor results hold up in production, Opus 5's defining feature will not be one benchmark record. It will be the ability to finish difficult work with fewer corrections at a cost that teams can justify for regular use.
(Photo: Ged Mullen-Buick / Unsplash, license.)
