AI Intelligence

Fable Matches Opus More Cheaply

7 stories · ~7 min read

Fable Matches Opus More Cheaply

Listen

If You Only Read One Thing

The expensive model can lower the bill; the cheap model can raise it. Fable 5.2 matches Opus coding accuracy more economically in Anthropic's testing, while Plotly's Muse Ember trial costs eight times Luna for the same score. These launches change which models deserve a comparison. They also make the amount of work an agent chooses to do harder to ignore.

Fable Matches Opus Coding at Half the Reported Cost

Fable 5.2 makes a stronger case for replacing expensive reasoning with a better model. In Anthropic's early coding test, Fable at High effort scored 45.3%, essentially matching Opus 5.5 at Max's 45.5%, for roughly half the cost per task. The test, FrontierCode 2.0, judges whether repository changes are suitable to merge.

That changes the comparison from “which model has cheaper tokens?” to “which setting buys this standard of work?” Anthropic's November 9 release prices Fable at $8 per million input tokens and $40 per million output tokens. Fable and the restricted-access Mythos 5.2 share an underlying model; Fable adds cyber and biology safeguards. General coding access is available now.

Independent results confirm a capability advance without confirming that particular coding discount. Artificial Analysis puts Fable's Max setting first at 65 on its broad Intelligence Index. It also reports $26.54 per index task. That's a different collection of tasks from Anthropic's coding comparison, so combining the two into a universal savings claim would be wrong.

The upgrade also changes how instructions affect the work. Anthropic's model-specific prompting experiments found that telling Fable it owned a codebase made it consolidate repeated code in about half of small-task trials, versus fewer than one in ten without that wording. The requested feature worked in nearly every trial either way.

Ownership wording bought extra renovation, rather than a demonstrated improvement in feature completion. Output grew about 50%, and the model changed existing feature behavior without being asked in roughly one trial in six to ten. Even an earlier instruction against incidental refactoring didn't fully cancel the effect.

That's a reason to make authority concrete: permission to repair a defect and permission to redesign surrounding code are different assignments. A general instruction to “take ownership” can quietly enlarge the second one. The published prompt makes this behavior reproducible enough to investigate, though the small vendor-run tasks don't settle its effect on production code.

Data terms can rule out the comparison altogether. Anthropic retains usage for safety monitoring for 30 days; eligible enterprise customers can keep zero-retention access, with customer-hosted monitoring in private beta.

Saturday's reasoning comparison held Opus fixed. This launch changes the model itself. On retained repository tasks, the telling result would be Fable High preserving Opus Max's accepted changes while lowering the full bill, including review of unrequested edits.

Muse Ember's Cheap Tokens Don't Guarantee Cheap Analytics

Meta's Muse Ember makes the low-price model tier more capable, but it already has a concrete counterexample to the claim that it will be cheaper for every agent job. Plotly got the same analytics score from Ember and GPT-6 Luna, with an eightfold difference in run cost.

Meta released Ember on November 9, targeting repeated agent work such as extraction, tool calls and coding subtasks. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. The model is available through the Meta Model API and Muse Code; its weights aren't public. Above that prompt threshold, both rates triple.

Artificial Analysis's independent evaluation gives Ember at Max an Intelligence Index score of 36 and a cost of $0.14 per task. That's credible evidence for considering it alongside other inexpensive workers. The index combines several kinds of work, however, and an average cannot identify the cheapest model for every one of them.

Plotly supplies the workload-specific check. Its 43-question analytics exam gave Ember and Luna 38 correct answers each. Ember cost $1.04 against Luna's $0.13. Plotly ran Ember through OpenRouter using its existing harness and grading, on private synthetic wind-fleet data designed to avoid a publicly available answer key.

Ember processed about 26.6 million tokens over the exam, nearly five times Luna's volume. An agent can repeatedly inspect tables, query results and earlier context before producing its final answer. Each additional step can be cheap while the complete investigation becomes expensive. Cached-token discounts also mean that multiplying raw input volume by the headline uncached rate won't reconstruct the bill.

Plotly's run doesn't make Ember a bad model. Its aggregate benchmark position remains attractive, and analytics is only one possible assignment. The comparison is also a single run per model, without evidence here that the two providers used identical internal serving settings. It supports a narrower conclusion: equal token prices and similar capability headlines weren't enough to predict this application's cost.

For model selection, that makes the repeated investigation the unit worth retaining: the same questions, tools, grading and complete run expense. The cheaper worker needs to stay cheaper after choosing how much to read. Ember would reverse Plotly's result if a repeated matched exam preserved 38 correct answers while bringing its bill down to Luna's published $0.13.

The Contrarian Take

Everyone says: A cheap model that approaches a premium model's benchmark score is the obvious worker for routine agent tasks.

Here's why that's wrong: “Routine” describes our expectation of the assignment, not how much work the model will perform. Ember and Luna reached the same score in Plotly's test, yet the bills differed eightfold. Fable's ownership experiment shows another route to extra expense: a prompt can purchase unsolicited renovation without changing whether the requested feature works. Model routing that considers only difficulty and token price misses how each model interprets the job. The economical worker is the one that reaches the required result with an appropriate amount of investigation and change.

Under the Radar

  • Ripple brings another coding model to local machines. Ollama's November 9 release supports Ricursive's Apache-2.0 Ripple preview, a 65-billion-parameter dense model. Its downloadable variants occupy roughly 39–46GB before runtime overhead, so this is a substantial local deployment. A dense model uses all its parameters for each token; a similar parameter count in a sparsely activated model can mean very different compute needs. Local availability gives private coding experiments another candidate, without proving interactive speed or reliable repairs. Model and runnable entry point.

  • Claude Code repairs credential redirects and permission checks. The 2.1.321 changelog says MCP login redirects could send secrets, including refresh tokens, to another host. It also fixes cases involving unprompted Bash execution outside the sandbox and permission checks on copy or move destinations. These are execution and credential-routing repairs, distinct from yesterday's saved-permission-file leak. The notes identify corrected paths, without measuring how often users encountered them. Primary changelog.

Quick Takes

  • Fireworks discloses another intrusion and suspends dataset operations. Its November 9 bulletin describes an October 24 intrusion through a cloud provider, alongside wider exposure from the earlier breach. Reported impact includes billing data and, for some accounts, secrets, datasets and incomplete model weights. Dataset uploads and downloads are suspended; Fireworks hasn't detected impact to its zero-retention Inference, Training SDK or Nexus services. Stored training assets and transient inference can have different security outcomes under one account. (Source)

  • Google adds confidential inference hardware; NEAR shows its limit. Google's TPU announcement adds 8tx for confidential inference and 8i Max for larger contexts. NEAR's November 9 incident report says an SGLang flaw let an attacker execute code inside an inference server and access requests, including confidential ones. Hardware isolation protects against the host; it doesn't repair vulnerable software inside. NEAR says it patched all servers. (Source)

  • Ollama stops silently shortening Responses requests by default. Version 0.40.16 respects the request's truncation setting. It also reports about 27GiB lower memory use for Qwen3.8 Flash Next on Apple Silicon. The request change makes an oversized input's treatment explicit, rather than allowing an answer that quietly lacked part of the supplied context. The memory figure is a release-note claim, not a measured improvement in coding quality or completion time. (Source)

The Thread

Model upgrades are making the spending target harder to hold still. Fable can reach an existing coding standard with less expensive reasoning, but a broader assignment can consume the saving. Ember makes capable workers inexpensive to call, while an application's investigation pattern can still produce a larger bill. That creates two legitimate, different purchases: holding output quality fixed and spending less, or holding spending fixed and asking for better work. Without choosing between them, an upgrade can look disappointing even when it succeeds. A team that gets more thorough investigation for the same budget hasn't cut its bill; a team that gets the same accepted result for less hasn't necessarily raised quality. Release comparisons become more informative when they say which purchase the experiment attempted. Otherwise suppliers can advertise efficiency while customers silently spend it on ambition.

Predictions

  • I predict: On December 10, Muse Ember will still have no publicly downloadable Meta-released weights, despite its availability through hosted APIs. Meta launched it as a proprietary service; cheaper access doesn't imply a forthcoming local release. (Confidence: medium; Check by: 2026-12-10)

Issue date: November 10, 2026. Generated at 03:57 AM ET.

Tomorrow morning in your inbox.

Subscribe for free. 10-minute read, every weekday.