Claude Opus 5 is the first Opus model in Anthropic’s fifth generation, and on most published benchmarks it outperforms Claude Fable 5.
In this article, you’ll learn:
Chat with Opus 5 on Overchat AI.
Claude Opus 5 is Anthropic's new frontier AI model designed for complex agentic coding and enterprise work. Opus 5 achieves 96.0% on SWE-bench Verified, a benchmark where the model has to solve real code issues from GitHub, meaning that it solved 96% of the problems it was given.
Claude Opus 5 is a large language model from Anthropic, released on July 24, 2026 as the first Opus model in the Claude 5 generation. Anthropic positions it for complex agentic coding and enterprise work, and it replaces Claude Opus 4.8.
Main Opus 5 specifications:
Anthropic describes Claude Opus 5 as a step-change over Opus 4.8. The biggest gains are in:
You can access Claude Opus 5 on:
Opus 5 is also the new default model on the Claude Max plan and the best model available on Claude Pro.
Claude Opus 5 once again has refreshed most public model leaderboards. On coding and agentic benchmarks, it outperforms Claude Fable 5 across multiple benchmarks, while costing half as much per token. The scores below come from DataCamp's Claude Opus 5 analysis.
| Benchmark | Claude Opus 5 | Claude Fable 5 | Claude Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|
| SWE-bench Verified | 96.0% | 95.0% | 88.6% | — |
| SWE-bench Pro | 79.2% | 80.3% | 69.2% | 64.6% |
| Frontier-Bench v0.1 | 43.3% | 33.7% | 18.7% | 37.5% |
| ARC-AGI-3 | 30.2% | — | 1.5% | 7.78% |
| BrowseComp | 90.8% | — | — | 90.4% |
| OSWorld 2.0 | 70.6% | — | — | 62.6% |
Claude Opus 5 scores 96.0% on SWE-bench Verified, and on Frontier-Bench v0.1, Anthropic's own agentic coding evaluation, it scores 43.3% against 18.7% for Opus 4.8.
| Benchmark | Claude Opus 5 | Claude Opus 4.8 | Claude Fable 5 |
|---|---|---|---|
| SWE-bench Verified | 96.0% | 88.6% | 95.0% |
| SWE-bench Pro | 79.2% | 69.2% | 80.3% |
| Frontier-Bench v0.1 | 43.3% | 18.7% | 33.7% |
According to Anthropic, Claude Opus 5 completes larger refactors and end-to-end features without leaving stubs or placeholders behind, and it performs best when given the full task specification up front rather than being steered turn by turn.
Anthropic also reports high precision and recall on code review, with Claude Opus 5 finding real bugs at a high rate per pass and producing few false positives, even at lower effort settings.
ARC-AGI-3 is designed to test problems a model cannot have memorized. Simon Willison, who covered the launch, described Claude Opus 5 as "relentlessly proactive" and cited an example where the model, asked to rebuild a machine part from a drawing it could not directly access, wrote its own computer vision pipeline to extract the geometry from raw pixels.
This is reflected in the scores. Claude Opus 5 scores 30.2% on ARC-AGI-3, against 1.5% for Opus 4.8 and 7.78% for GPT-5.6 Sol. Anthropic reports the result as roughly 3x the next-best model.
| Benchmark | Claude Opus 5 | Claude Opus 4.8 |
|---|---|---|
| ARC-AGI-3 | 30.2% | 1.5% |
Anthropic built Claude Opus 5 for long-horizon agentic work, and states that its instruction following and tool calling stay consistent across the full 1M token window. Anthropic also reports around 1.5x the next-best model on Zapier's AutomationBench.
| Benchmark | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|
| OSWorld 2.0 | 70.6% | 62.6% |
| BrowseComp | 90.8% | 90.4% |
Devin CEO Scott Wu said that on FrontierCode 1.1, "Claude Opus 5 approaches Fable-level performance at half the cost."
Claude Opus 5 changes several defaults that Opus 4.8 users will notice. Here's what you need to know:
Claude Opus 5 has thinking on by default. On Opus 4.8, requests ran without thinking unless you set thinking: {"type": "adaptive"}. On Claude Opus 5, thinking is engaged from the start. Because max_tokens caps total output including thinking, Anthropic recommends revisiting that value on any workload carried over from Opus 4.8.
Disabling thinking on Claude Opus 5 requires effort high or below. Setting thinking: {"type": "disabled"} at xhigh or max returns a 400 error.
The effort parameter has a larger impact on Claude Opus 5 than on earlier Opus models. According to Anthropic, Opus 5 users should start at the high default and adjust in both directions, because low and medium produce strong quality at a fraction of the tokens.
Claude Opus 5 caches prompts from 512 tokens, down from 1,024 on Opus 4.8.
Claude Opus 5 supports mid-conversation tool changes in beta. With the mid-conversation-tool-changes-2026-07-01 header, you can add or remove tools between turns while preserving the prompt cache.
Claude Opus 5 has a fast mode on the Claude API, running 2.5x faster at double the base rate — $10 per million input tokens and $50 per million output tokens. This is the lever to pull when latency matters more than cost.
Anthropic documents several behavior changes worth knowing before migrating a prompt:
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens on the Claude API. The cost is the same as Claude Opus 4.8.
| Model | Input per MTok | Output per MTok |
|---|---|---|
| Claude Opus 5 | $5 | $25 |
| Claude Opus 5 (fast mode) | $10 | $50 |
| Claude Fable 5 | $10 | $50 |
| Claude Opus 4.8 | $5 | $25 |
| Claude Sonnet 5 | $3 | $15 |
| Claude Haiku 4.5 | $1 | $5 |
You can also access Opus 5 on Overchat AI, an all-in-one AI app we've built, where it is included inside the Pro subscription. Overchat AI Pro pricing is as follows:
| Pro Plan | Monthly | Annual |
|---|---|---|
| Pro | $14.99/month | $49.99/year |
After the release of Claude Opus 5, there are very few practical reasons to use Fable, since Opus 5 outperforms it while costing less. This is true for coding and reasoning tasks. One exception, where Fable clearly has an advantage over Opus (forgive the pun), is cybersecurity and offensive testing.
If you're working on very difficult coding problems, you may have a slightly higher success rate with Claude Fable 5 — it has a narrow lead on SWE-bench Pro (80.3% vs 79.2%).
Claude Opus 5 is behind Mythos 5 on offensive cybersecurity. Anthropic reports it approaches Mythos 5 on identifying vulnerabilities but stays substantially behind on exploiting them, because it was not trained on weaponization.
Claude Opus 5 produces longer output than Opus 4.8. Default responses and written deliverables both run longer, and the model narrates its progress during agentic sessions.
Claude Opus 5 expands the scope of a task. Anthropic notes that the model adds steps that were not requested and applies its own judgment about what a task should be — you might find that you need to start adding explicit constraints to prompts to control the scope.
Running Claude Opus 5 with thinking disabled produces artifacts. The model occasionally writes a tool call into its visible text instead of emitting a structured tool_use block, and the call never runs. In an agentic loop, the leaked text stays in the conversation history and affects later turns. It can also emit thinking tags into the visible response. Anthropic recommends keeping thinking on and controlling cost with lower effort levels instead.
Claude Opus 5 is available on Overchat AI now, in the model picker on both desktop and mobile.
Overchat AI is an all-in-one platform, so you can switch from Claude Opus 5 to other models, including Fable 5, GPT-5.6, Gemini, Kimi K3, and more. The context is shared between them, so you can reference previous chats when talking to a new model.
Common questions about Claude Opus 5, its pricing, and how it compares to other Claude models.
Claude Opus 5 outperforms Claude Fable 5 on multiple benchmarks, including Frontier-Bench (43.3% to 33.7%), ARC-AGI-3, BrowseComp, and OSWorld 2.0, at half the price per token.
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens on the Claude API, the same rate as Claude Opus 4.8. Fast mode doubles that to $10 and $50. Through Overchat AI, it's included in the $14.99/month Pro plan.
Claude Opus 5 has a 1M token context window, which is both the default and the maximum. Max output is 128k tokens, or up to 300k through the Message Batches API with a beta header.
Claude Opus 5 requires a paid plan on Anthropic's own subscriptions, where it is the default model on Claude Max and the strongest model on Claude Pro. On Overchat AI, Opus 5 is included in the Pro plan at $14.99/month or $49.99/year.
Claude Opus 5 has a reliable knowledge cutoff of May 2026. Fable 5 and Sonnet 5 both have knowledge cutoff dates of January 2026.
Claude Opus 5 is available on Claude.ai, Claude Code, Claude Cowork, the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. It's also on Overchat AI, which bundles it with the other frontier models under one subscription.
At the time of writing this, Claude Opus 5 is the best Claude model for coding and long-running agentic work. Yes, it even outperforms Fable, and given its lower price per token, it's the new recommended default for developers and vibecoders. And if you're wondering where to access Claude Opus 5, the model is already available on Overchat AI.
