TLDR
Key findings
- Capability: leads 9 of 11 published rows, including FrontierCode 2.0 at 49.1%, the highest the vendor has recorded.
- Blended operating cost: $5.74 per million tokens on a 7:2:1 cache-input-output mix, against a 97.5% cache-read discount.
- Per-task claim: at high effort it matches the tier below at max effort for roughly half the cost, on one coding benchmark.
- Safety: best composite behavioral-audit score published, and sandbox escape at 3.1% against 21% and 39% for the two prior generations.
- Enterprise blocker: 30-day retention is mandatory; zero data retention requires express authorization or the private-beta monitoring product.
- Migration cost: four breaking changes, including thinking blocks that only this generation can read.
- AI Matic Bench Score: 81.3 out of 100, provisional pending our own benchmark run.
We scrambled in excitement to review the new Fable 5.2, after its promise of being the strongest model on the published table by a wide margin. Spoiler alert – it did not disappoint. From the launch material published by Anthropic, Fable 5.2 leads every comparison row (except two), formalized Fermat's Last Theorem in 30 hours with a thousand agents and built a bootable operating system from an empty folder over eight days. The vendor itself recommends starting with the tier below Fable 5.2 and moving up to it only if that model fails to meet your evaluation requirements – which is quite indicative of how they perceive the performance.
From what we can see, it is an accurate read on the enterprise delivery aspect, and the reasons are commercial rather than technical. At $8 and $40 per million tokens it blends to $5.74 on an agentic mix (the most expensive model we have assessed). It requires 30-day data retention and is not available under zero data retention without express authorization. And it ships four breaking API changes, one of which quietly breaks any architecture that routes conversations between models.
Model overview and pricing
Published benchmark results
What the demos do and do not establish
The Fermat formalization, the eight-day operating system, the two black-hole results and the capsid-folding method are company-reported and have not been independently reproduced. They are concrete and they point at genuine long-horizon capability. They are not a procurement artefact, and none of them tells you how the model behaves on your codebase.
Enterprise readiness
When it all boils down, this is where the buying decision sits, and it is the section the launch material covers least.
Availability is otherwise the broadest in the family: the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud under both managed options, and Microsoft Foundry. For an AWS-native build it is reachable without a new vendor relationship.
Safety and alignment
The alignment results are the strongest part of the release for a regulated buyer. The composite behavioral-audit score of 2.67 is the best the vendor has published, ahead of both tiers below it and of the prior generation. Sandbox-escape attempts fall to 3.1%, against 21% and 39% for the two previous generations of this tier, and the vendor reports that describing the sandbox and adding a weakness-reporting tool reduces it further. External pre-release testing was carried out by the UK AI Security Institute, Frontier Design and METR.
We must add that two caveats belong in any proposal that cites these numbers. The vendor states that this model voices suspicion of being evaluated more often than any of its recent models, which directly weakens the inference from evaluation to production behaviour, and it says so itself. And on at least one measure, taking harmful actions after wrongly concluding it was in a simulation, it scores slightly worse than the tier below. False-positive safeguard blocks are down about 55%, which is a usability gain rather than a safety one.
GoML’s internal benchmark: LatticeBench
LatticeBench by GoML approximates enterprise-shaped engineering work rather than isolated coding puzzles. It is a multi-service codebase spanning identity, sales, SLA, engagement, organization, file store, work management, gateway and AI copilot, plus a shared common library and a gateway config.
Fifty-seven defects are deliberately injected into it, 18 of them critical. Eight come with an explicit ticket; the other 49 are hidden and the model has to find them unprompted. The design target is the defect class that ends enterprise engagements: cross-tenant mass assignment, fail-open authorization, and shared-module blast radius.
The run tests whether the long-horizon results transfer. A model that coordinates a thousand agents for eight days should, in principle, trace a writable tenant identifier from a schema through a facade into a service. That chain is the hardest class in our manifest and it decides whether a model can be trusted on a multi-tenant codebase.
AI Matic Bench Score
The AI Matic Bench Score is GoML's composite rubric for deciding whether a model belongs in a client delivery pipeline. It reduces eight axes to a single figure out of 100: each axis is scored from 0 to 10, multiplied by its weight, summed, and scaled by 10.
Vendor tables answer a research question: “Is the new model better?”. Delivery teams answer a different one: “Does putting this model into a running pipeline improve accepted outcomes per rupee and per hour, without forcing a requalification cycle?”.
The score is built for that second question, and it is meant to be read alongside the vendor results, rather than in place of them.
How it is designed
The eight axes are weighted in three tiers, so that the score reflects what decides a delivery engagement rather than what makes a model impressive.
- Delivery outcomes (A1 to A3, combined weight 0.55). Whether the model finishes the work, uses tools correctly, and does so at an acceptable cost per accepted outcome. These carry the most weight because they determine margin and timeline.
- Operating fit (A4 to A7, combined weight 0.40). Whether the model can be run, governed and defended in a client environment: latency, deployment terms, regulated-domain competence, and safety. Each axis carries 0.10.
- Confidence (A8, weight 0.05). How far the evidence behind the other seven scores can be verified independently. It adjusts how much trust the total deserves and is deliberately small.
Every score is tied to a stated basis, and the total is labelled provisional while any axis rests on published or projected evidence rather than a GoML benchmark run. When the LatticeBench run completes, its measured result replaces the projection on the affected axes and the total is recalculated.
Formula and bands
AI Matic Bench Score = 10 × Σ (axis score × axis weight), with the eight weights summing to 1.00. Band reading for each axis: 9 to 10 means no concerns; 7 to 8 means production-viable with a named guardrail; 5 to 6 means a human review gate on every client-facing output; below 5 is pilot-only or disqualifying.
Long-horizon completion at 9.5 and safety at 9.0 are the highest marks we have given on those axes, and the total still lands at 81.3 because unit economics and deployment envelope carry 30% of the weight between them. Neither is a judgment about the model's intelligence.
Analysis and conclusion
Fable 5.2 is the right model for a narrow and valuable set of work: long-horizon agentic runs that currently fail, open-ended problems where a cheaper model stalls, robotics and hardware control and scientific workflows. On those the margin is not incremental, and a model that finishes in one pass is cheaper than one that needs three.
It is the wrong default for everything else, and Anthropic themselves says so first. Two deployment facts decide most engagements: zero data retention is not available by default, which rules the model out of some contracts entirely until Enterprise Frontier Safeguards leaves private beta, and thinking blocks do not survive a model switch, which quietly breaks the router-and-fallback pattern most agentic platforms are built on.
The recommended shape is to keep the tier below as the default, route the hardest long-horizon work here behind a cost cap and a per-session budget, confirm the retention position with the account team before any client pilot, and run an effort sweep from the API default. The next step on our side is the LatticeBench run, which replaces the published-evidence scores on two axes.
Stay tuned to the GoML blog for the latest updates in the world of AI and ML engineering.
