Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max)
Anthropic’s new Claude Opus 5.5 model is drawing attention for offering roughly double the speed and half the per-task cost of Opus 5 at high reasoning levels, while still ranking among the most capable proprietary LLMs. Commenters question the real-world usefulness of the extremely expensive “max” reasoning mode, reporting cases where the model overthinks tasks until it hits token limits and noting that medium or high settings often deliver better value. Broader concerns surface around benchmark relevance, potential performance regressions over time, and whether top-tier proprietary models can justify their prices against increasingly competitive open-weight alternatives.
Pricing, Cost, and Performance
- Opus 5.5 Max is about half the cost per task and roughly twice as fast as Opus 5 at similar “high” effort, which many see as a strong improvement.
- Some note Astra High is still slightly cheaper than Opus 5.5 High; others argue “value” models like Luna offer far better cost-per-work than frontier models.
- Confusion and criticism over wording like “somewhat expensive compared to models of similar price” and “somewhat expensive when comparing to other models of similar price,” which is seen as meaningless without a performance anchor.
Reasoning Modes and ‘Max’ Overthinking
- Several users report Max/xhigh reasoning often overthinks, hits token or tool limits, or never returns an answer (e.g., failing “SVG pelican on a bicycle” after burning through the entire thinking budget).
- Medium or High reasoning is widely seen as the practical sweet spot; Max is viewed by some as “benchmaxxing” only, not for real work.
- Higher effort is reported to increase verbosity and self-referential loops rather than quality, especially in agentic or long-running tasks.
Benchmarks, Evals, and Model Degradation
- Many distrust static benchmarks: models may be tuned to ace them, they rarely reflect real usage, and they’re often run under idealized settings.
- Calls for continuous, randomized evals over time, including degradation trackers; one commenter reports Sol regressing to Luna-level performance in internal tests.
- Some suspect providers could “pull the rug” post-launch; others (including one identified as working at a vendor) attribute perceived regressions mostly to randomness and rising expectations.
User Experiences and Workflows
- Several users preferred Opus 4.8 over 5 due to better instruction following and task continuity; they hope 5.5 fixes this.
- Medium reasoning is commonly used for interactive coding and review; high/Max is seen as wasteful and unstable.
- Reports that some models overengineer solutions, especially for infra or code tasks; users try to constrain tool usage or reasoning effort.
- Debate over agentic workflows: some successfully run long-lived agents (days) to fix bug lists; others see long autonomy as unreliable for ill-defined real-world projects.
Broader Reflections
- Some feel progress is slowing and models are near a capability ceiling; others point to ongoing but incremental gains.
- There’s recurring tension between “best but expensive” closed models and “good enough” cheaper or open-weight models, with predictions that “good enough” may dominate many use cases.