Claude Platform and API Digest

Analysis

Read release notes as contracts

Platform release notes describe behavior that production clients depend on: context length, stop reasons, refusal semantics, effort defaults, cache thresholds, deprecations, and tool-use behavior. A weekly review should convert each note into a migration task, an eval update, or a monitoring change.

The two anchors for this issue are Opus 5 on July 24 and Fable 5.1 on September 1. Opus 5 is the recommended default; Fable 5.1 is the escalation path. Both changed the request and response surface in ways that will not show up as a clean deploy-time error.

Analysis

Opus 5: fallbacks and cache-preserving tool changes

Two Opus 5 changes matter to API clients. The beta for mid-conversation tool changes without invalidating the prompt cache removes a long-standing cost penalty for dynamic tool sets — previously, changing the tool list mid-conversation meant re-paying for the cached prefix. If your agent adds or removes tools as a task progresses, this is worth testing.

The second is automatic fallbacks when safety classifiers flag Opus 5 or Fable 5. Biology-related Fable blocks now route to Opus 5 rather than Opus 4.8. Your telemetry should be able to tell you which model actually answered, because "the model I requested" and "the model that responded" can now differ for classifier reasons.

• Log the responding model ID, not just the requested one.
• Test dynamic tool sets against the cache-preserving beta before assuming the old cost model.
• Separate classifier fallback from error handling in dashboards.
• Note the published cyber policy: finding vulnerabilities in source is permitted; binary scanning, pentest, and exploit generation are blocked.

Analysis

Fable 5.1: three breaking, five additive

The Fable 5.1 platform overview lists three breaking behaviors relative to Fable 5: forced tool use errors, earlier models cannot read its thinking blocks, and editing an earlier turn invalidates thinking. Each of these tends to appear as an intermittent failure in a specific code path rather than a startup error, so they need deliberate tests.

The additive betas are per-message effort, turn-scoped system messages, a display: "updates" progress mode, the lower cache-read rate, and content provenance. Per-message effort is the most architecturally interesting: it lets one conversation mix cheap and expensive turns, which is a cheaper lever than splitting a workload across two models.

Analysis

Cost accounting after the September changes

Three pricing facts changed or were clarified since July. Sonnet 5 stayed at $2 / $10 instead of rising to $3 / $15. Fable 5.1 and Mythos 5.1 cache reads bill at 2.5% of base input rather than 10%. Opus 5 held the Opus 4.8 rate of $5 / $25, with fast mode at roughly 2.5x speed for 2x base price.

Together these change routing math more than any capability claim. A workload with a high cached-read share may now be cheaper on the flagship than a naive base-rate comparison suggests, and a high-volume workload on Sonnet 5 is cheaper than the July forecast predicted.

• Recompute cost per successful answer per model, including cache reads.
• Reconcile any dashboard still assuming Sonnet 5 at $3 / $15.
• Treat fast mode as an explicit latency purchase, not a default.
• Keep the Batch API 50% discount in scope for non-interactive work.

Analysis

Long context is not free context

Fable 5.1, Opus 5, and Sonnet 5 all list 1M token context windows, but capacity does not remove retrieval design. Longer prompts increase cost, latency, and evaluation complexity. Use long context where preserving full source material matters; keep retrieval and chunking when the task needs a narrow slice.

The honest benchmark is answer quality per dollar at your required latency. In many systems a smaller context with good retrieval beats a huge context filled with unfocused documents.

Analysis

Deprecation hygiene

Published retirement floors give you a planning horizon: Haiku 4.5 no sooner than October 15, 2026, Sonnet 5 no sooner than June 30, 2027, Opus 5 no sooner than July 24, 2027, and Fable 5.1 no sooner than September 1, 2027. Haiku 4.5 is the nearest and is frequently embedded in high-volume paths, which is a bad combination to discover during an incident.