Claude Code Experiment Quietly Remapped Effort Levels for Fable 5, Anthropic Confirms

A developer spent an afternoon convinced their code was broken. They checked the t3 framework, audited their application logic, and began to wonder if something had gone wrong with their Mac. When they finally pulled up the raw API request logs from Claude Code, they found the answer: a single number — 10 — where they expected to see something much higher. They had selected "high" effort, the top named tier for Claude Fable 5 in Claude Code's model picker. The number in the logs was the same one that had historically corresponded to "low."

The discovery, posted to X by developer @argofowl on August 22, prompted Thariq Shihipar, a member of Anthropic's Claude Code team, to acknowledge publicly that the company had been running a server-side A/B experiment that changed how effort selections are translated into numeric values before reaching the inference layer. The experiment affected Fable 5 sessions on Claude Code version 2.1.236 and later. Users on older versions, and those working with Claude Opus 5, were not enrolled.

What the Effort Parameter Actually Does — and Why This Matters

The effort parameter — available as output_config.effort in Fable 5's API — is now the only inference-time control users have over reasoning depth. Unlike earlier Claude models, which exposed budget_tokens and temperature settings that developers could tune directly, Fable 5 eliminated those levers. Thinking cannot be disabled. Temperature cannot be set. The named effort tier — low, medium, high, xhigh, or max — is the sole dial.

Server-side, Anthropic maps each named tier to a numeric score before the inference engine executes the request. The experiment argofowl discovered changed that mapping: a request submitted as "high" arrived at the inference layer labeled 10 — the value that, in the prior implementation, corresponded to "low."

Shihipar's response was technically precise. According to his posts on X, the numeric scale itself changed in the experiment, meaning a "10" under the new scale is not the same as a "10" under the old one. The labeled tier still corresponds to the same point on the model's compute curve, he argued, and his team ran internal evaluations confirming no performance impact.

What developers cannot do is verify this from the outside. The API log is the most objective signal available. When that number matches the value associated with a lower tier, distinguishing a genuine behavioral change from a benign numeric remap requires either trusting Anthropic's word or conducting exactly the kind of afternoon-long investigation argofowl ran.

A Pattern Predating This Incident

The effort experiment did not appear in the Claude Code changelog. In April 2026, Anthropic published a detailed engineering postmortem acknowledging three separate Claude Code quality incidents from the preceding six weeks — including one that was directly analogous. On March 4, the company had changed Claude Code's default reasoning effort from high to medium to reduce latency, without marking it as a material behavioral change. Users noticed, complained, and Anthropic reverted the change on April 7, describing the original decision as "the wrong tradeoff."

The April postmortem committed to broader evaluation suites, gradual rollouts, and tighter controls on prompt changes. The current effort remapping happened three months later, again without a changelog entry.

A separate episode in late June 2026 saw Anthropic acknowledge that Claude Code had been applying hidden steganographic modifications to requests to detect unauthorized model distillation by competitors — also undisclosed until a developer discovered it through inspection.

Each incident involved a server-side change to how Claude Code actually behaved that was not reflected in the changelog. The pattern describes a company running significant configuration tests on production users before deciding whether to ship them — standard engineering practice, but one that creates a structural trust problem when the "configuration" being tested governs a developer's entire coding workflow.

Opus 5: A Separate and Larger Instability Problem

The effort experiment landed inside a sustained wave of criticism about Claude Opus 5, released July 24, 2026. Anthropic launched it with strong benchmark scores at roughly half the per-token cost of Fable 5, but the reception from practitioners has been substantially less positive.

Developers report a model that treats minor tasks as high-severity problems, produces sprawling responses to narrow questions, and takes unsolicited action beyond a prompt's scope. Dan Shipper, co-founder and CEO of Every, wrote on launch day that Opus 5 "argued with instructions, stopped before the work was finished, and generally didn't play well with our existing skills and plugins." A prominent venture capitalist publicly called it nearly unusable for real-world debugging. One developer filed GitHub issue #84002 on the Claude Code repository — a thousands-of-words technical account of being unable to keep Opus 5 on task across a multi-agent workflow — concluding that the root cause required Anthropic to fix something the user could not reach, and ultimately canceling multiple Claude subscription plans.

Anthropic's status page recorded separate degraded performance incidents for Claude Opus 5 on August 5 and again on August 17 and 18 — infrastructure failures distinct from the behavioral complaints but reinforcing the perception of an unstable release.

Shihipar acknowledged the issue. "Opus 5 is a really spiky model," he wrote in response to user criticism on X, "and we want our models to be consistent and warm and feel like Claude. We're working hard on this and it is a huge priority for us." The company has not published a timeline for addressing it.

The Infrastructure Trust Problem AI Providers Haven't Solved

Both stories point to a structural gap that is not unique to Anthropic. Benchmarks — SWE-bench, Terminal-Bench, ARC-AGI-3 — measure bounded tasks under controlled conditions. They say nothing about whether a model will honor intent consistently on a 40-minute autonomous coding run, or whether its effort level will mean the same thing it meant last week. Daily developer work requires the kind of behavioral invariance that benchmarks are not designed to test.

In traditional software, a changelog makes the answer to "what changed?" discoverable. Claude Code has a changelog, but it does not cover service-layer configuration experiments. The model a developer selects in the picker is not a fixed artifact; it is a named pointer to server-side infrastructure that Anthropic can reconfigure per session, per traffic segment, without client-side visibility.

As Claude Code becomes infrastructure for more production workflows, the changelog gap becomes a liability. The April postmortem showed Anthropic understands this and has started to address it — the commitments it made were real. The current episode shows those commitments were drawn around a narrower set of changes than the ones that matter to developers. Whether the effort experiment becomes a permanent scale change or is rolled back, the more durable question is whether Anthropic extends changelog coverage to the server-side configuration layer — the place where, right now, observable behavior can change without leaving any trace a developer can find.