Anthropic revives Opus with version 5, claiming near-flagship performance at half the price
The refreshed model—codenamed Honeycomb—ships with a 1M-token context window and xhigh reasoning mode, closing most of the gap with Fable 5 while undercutting it on cost.
What matters
- Opus 5 rolled out July 23, 2026, succeeding Opus 4.8 and approaching Fable 5 performance at half the price.
- The model ships with a 1M-token context window, an xhigh reasoning effort mode, per-turn controls, and a safety fallback to Opus 4.8.
- Opus 5 was codenamed Honeycomb; an early-access preview briefly appeared in Cursor's model picker on July 9, 2026 before being pulled.
- Anthropic says Opus 5 is better at knowledge work than any prior model, substantially better at coding than Opus 4.8, and exhibits its lowest rates of deceptive behavior.
- A full public model card and pricing page had not yet been published at the time of reporting; Opus 4.8 was priced at $5/M input and $25/M output.
Launch facts
- Price:
- Half the price of Fable 5 (exact per-token pricing not yet publicly confirmed); predecessor Opus 4.8 was $5/M input, $25/M output
- Availability:
- Rolling out across providers starting July 23, 2026
- Platforms:
- Anthropic API, Supported providers (including Cursor)
What happened
Anthropic has released Opus 5, the latest iteration of its Opus model line and the successor to Opus 4.8, which had served as the Opus flagship since late May 2026. According to the company, Opus 5 delivers capabilities that come close to Fable 5—Anthropic's most powerful commercially available model—at half the price.
The model began rolling out across providers on July 23, 2026, though it first surfaced publicly earlier in the month. A research preview codenamed "Honeycomb EAP" briefly appeared inside Cursor's model picker around July 9, 2026, before being pulled within hours. The spec profile that leaked at that time—a 1 million token context window, an "xhigh" reasoning effort mode, per-turn controls, and a safety fallback to Opus 4.8—matches the model now shipping.
On performance, Anthropic says Opus 5 is better at knowledge work than any of its past models and substantially better at complex coding tasks than Opus 4.8. The company also claims improvements in scientific research, citing tasks like predicting how variations in a protein sequence might affect molecular function.
On safety, Anthropic says Opus 5 is more resistant to being tricked and "exhibits the lowest rates of deceptive behavior" of any model it has released. Notably, the company says it deliberately avoided training Opus 5 on cyber-related tasks. Despite that, the model's general capability improvements make it broadly better at cybersecurity tasks than its predecessor—though Anthropic acknowledges it remains "substantially behind" its top-tier model in that domain.
Anthropic's most powerful model, Mythos, continues to be available only to a limited number of vetted organizations through the company's Project Glasswing initiative.
For pricing context, Opus 4.8 was priced at $5 per million input tokens and $25 per million output tokens. A full public model card and pricing page for Opus 5 had not yet appeared at the time of reporting, and Anthropic has said little on the record about the rollout beyond its performance claims.
Why it matters
The Opus line had begun to feel neglected as Anthropic pushed Fable 5 and the restricted Mythos model. Opus 5's release signals that Anthropic sees a competitive mid-tier offering as essential—near-flagship quality at half the cost could make it the default choice for developers and enterprises that don't need the absolute top tier.
The 1M-token context window and xhigh reasoning mode bring Opus 5 into line with what power users increasingly expect from frontier models, while the deliberate exclusion of cyber-specific training is a notable safety posture. It suggests Anthropic is trying to balance broad capability with responsible deployment, even as the model's general improvements make it more useful for security work incidentally.
The lack of a public model card and pricing page at launch, however, means customers are largely operating on Anthropic's word. Independent benchmarks are not yet available.
What to watch
- Whether third-party benchmarks confirm that Opus 5 genuinely approaches Fable 5 quality, particularly on coding and research tasks.
- When Anthropic publishes a full model card and pricing page with exact per-token costs.
- How competitors like OpenAI and Google respond on pricing for their own mid-tier models.
- Whether the Honeycomb EAP leak in Cursor signals broader challenges for Anthropic in controlling pre-release information.
What to do next
Developers
Run your existing coding and knowledge-work benchmark suites against Opus 5, Opus 4.8, and Fable 5 to validate Anthropic's performance claims before migrating.
Vendor-reported gains need independent verification against your specific workloads, especially for complex coding tasks where 'substantially better' is a qualitative claim.
Founders
Model what a 50% price reduction for near-flagship performance means for your per-query unit economics, using Opus 4.8's $5/M input and $25/M output as a baseline.
If Opus 5 delivers comparable quality at half the cost of Fable 5, it could materially lower margins for AI-heavy products—but exact pricing is not yet confirmed publicly.
PMs
Pilot Opus 5 in a non-production environment for your highest-volume use cases, paying special attention to the 1M-token context window and xhigh reasoning mode.
The cost savings and expanded context are compelling, but you need to confirm that 'nearly matching' Fable 5 meets your quality bar before switching production traffic.
Investors
Assess how Anthropic's tiered model strategy—Opus 5 as mid-tier, Fable 5 as commercial flagship, Mythos as restricted access—positions the company against OpenAI and Google in the enterprise market.
A strong mid-tier offering at half the flagship price could expand Anthropic's addressable market and improve revenue diversification, but the lack of a public pricing page creates near-term uncertainty.
Operators
Review whether Opus 5's improved safety profile and lower deceptive behavior rates allow you to simplify output-filtering layers in your pipeline.
If the model is genuinely more resistant to manipulation, you may be able to reduce guardrail infrastructure overhead—but claims should be tested with adversarial prompts before relying on them.
How to test
- 1Obtain API access to Opus 5 through Anthropic's platform or a supported provider.
- 2Run the same prompt set against Opus 5, Opus 4.8, and Fable 5 (if available) to compare output quality on coding, knowledge work, and research tasks.
- 3Measure per-token cost for identical requests across the models to verify the claimed 50% savings relative to Fable 5.
- 4Test the 1M-token context window with long-context workloads to evaluate retrieval and coherence at scale.
- 5Evaluate the xhigh reasoning effort mode on complex multi-step problems and compare latency against standard reasoning settings.
- 6Run adversarial or jailbreak-style prompts to test Anthropic's claims about lower deceptive behavior and improved resistance to manipulation.
Caveats
- All performance claims are vendor-reported; independent benchmarks are not yet available.
- A full public model card and pricing page had not appeared at the time of reporting, so exact per-token costs for Opus 5 are unconfirmed.
- Fable 5 and Mythos may not be accessible to all accounts, limiting direct comparison.
- The kie.ai source is a secondary blog; its technical specs (1M context, xhigh mode) have not been independently confirmed by Anthropic's own documentation.
- Cybersecurity task performance may vary significantly from Anthropic's general claims depending on specific use cases.