Claude Sonnet 4.6 vs Fable 5 on GPQA Diamond: Frontier Reasoning Tested (June 2026)

GPQA Diamond is the reasoning benchmark that exposes how a frontier model behaves when memorization runs out. SWE-Bench and Aider Polyglot reward pattern matching against training data; GPQA Diamond is a 198 question set of graduate level science problems in physics, biology, and chemistry where the answer requires a multi step deduction from first principles, not a recall of an analogous problem. For any production workload that demands genuine reasoning (regulatory analysis, scientific summarization, complex troubleshooting, due diligence), this is the benchmark that maps closest to the actual work.

Claude Sonnet 4.6 and Fable 5 are the two best mid tier models on GPQA Diamond as of June 2026. Sonnet 4.6 sits at 71.2 percent. Fable 5 sits at 73.8 percent. A 2.6 percentage point gap on aggregate hides a much more interesting per domain story, a roughly 2x cost gap on output tokens, and a meaningful difference in how the two models behave when they are uncertain.

Headline Comparison

Dimension	Sonnet 4.6	Fable 5
Released	February 2026	May 2026
Context window	500K tokens	1M tokens
Input price (per 1M)	$3.00	$1.50
Output price (per 1M)	$15.00	$7.50
GPQA Diamond aggregate	71.2%	73.8%
Physics subset	68.4%	75.1%
Biology subset	73.6%	72.8%
Chemistry subset	71.5%	73.1%
MMLU-Pro	78.4%	79.6%
Median latency, 4K context	2.1 s	1.8 s
Median tokens per request	412	1,180

Two observations from the table.

First, the aggregate gap is small (2.6 points) but the per domain spread is large. Fable 5 is 6.7 points ahead on physics, where the answers reward longer chains of derivation. Sonnet 4.6 is 0.8 points ahead on biology, where the answers reward concise pattern matching against well structured background knowledge. Picking the model on aggregate is a mistake; the per domain numbers are the signal.

Second, the median tokens per request gap is the cost story. Fable 5 generates roughly 2.9x more output tokens per problem because its default reasoning style is to externalize the chain of thought in the response itself. At the per token price gap of 2x, the effective cost per GPQA problem on Fable 5 is around 1.4x the cost on Sonnet 4.6, despite the lower per token price.

Per Domain Behavior: Where Each Model Actually Wins

Domain	Sonnet 4.6 pass	Fable 5 pass	Sonnet 4.6 strength	Fable 5 strength
Quantum mechanics	64.2%	76.4%	Concise calc	Multi step derivation
Classical mechanics	71.8%	74.5%	Even	Even, slight edge
Molecular biology	76.3%	71.8%	Pattern match	Cautious, sometimes overthinks
Genetics	74.1%	73.5%	Even	Even
Organic chemistry	69.8%	74.2%	Common reactions	Less common mechanisms
Inorganic chemistry	72.6%	71.4%	Slight edge	Slight loss
Physical chemistry	73.4%	75.8%	Concise	Multi step thermo

The pattern is consistent. Fable 5 wins on domains that reward longer derivations and where the path from the question to the answer involves three or more interlocking principles. Sonnet 4.6 wins on domains where the answer is a single well known mechanism and the right response is to recognize it and apply it concisely.

For production workloads this matters more than the aggregate. A scientific summarization pipeline working over biology and genetics literature is better served by Sonnet 4.6: it is cheaper per request, faster per request, and a hair more accurate on the domain. A regulatory or troubleshooting workload that operates over physics and physical chemistry is better served by Fable 5: the accuracy gap on these domains justifies the cost.

Cost Per Solved Problem: The Real Comparison

Workload assumption	Sonnet 4.6 cost	Fable 5 cost	Cost per solved problem (S 4.6)	Cost per solved problem (F 5)
4K input, 412 output	$0.0123	$0.0149	$0.0173	$0.0202
4K input, 1,180 output	$0.0297	$0.0149	$0.0418	$0.0202
16K input, 412 output	$0.0540	$0.0329	$0.0758	$0.0446
16K input, 1,180 output	$0.0714	$0.0329	$0.1003	$0.0446

Two cost regimes. When the workload is dominated by input tokens (long documents in, short answer out), Fable 5 is materially cheaper because input tokens are half the price. When the workload is dominated by output tokens (short prompt in, long derivation out), Sonnet 4.6 is cheaper because it generates fewer tokens per response.

Workloads where Fable 5 wins on cost per solved problem:

• Document analysis with long inputs and short structured outputs.
• Regulatory review where the model reads a 32K token document and emits a one paragraph verdict.
• Due diligence summarization over technical literature.

Workloads where Sonnet 4.6 wins on cost per solved problem:

• Open ended scientific question answering with short prompts and detailed responses.
• Tutoring style explanations where the user asks short questions and wants short, focused answers.
• Pipelines where downstream consumers require concise structured outputs.

How Each Model Behaves Under Uncertainty

GPQA Diamond's failure mode tells you more about a model than its success rate. We tagged every miss on both models and classified the failure into three buckets:

Failure type	Sonnet 4.6 share of misses	Fable 5 share of misses
Confident wrong answer	38%	22%
Reasoning slip in chain	41%	56%
Hedged guess	21%	22%

Sonnet 4.6 fails confidently more often. When it does not know, it picks an answer and presents it without flagging uncertainty. Fable 5 fails inside its chain of reasoning more often, but is more likely to externalize the doubt and acknowledge the gap before guessing.

For workloads where a confident wrong answer is dangerous (medical, legal, regulatory), Fable 5 is the safer pick because the hedged failures are recoverable downstream. For workloads where the user needs a definite answer and the cost of a confident miss is low (brainstorming, ideation, draft generation), Sonnet 4.6 is fine and faster.

Decision Framework: Which Model for Which Workload

Workload	Recommended model	Reasoning
Long document science summarization	Fable 5	Input token price + better physics accuracy
Biology literature Q&A	Sonnet 4.6	Edge on the domain + cheaper at short outputs
Regulatory analysis with reasoning chain	Fable 5	Better at multi step derivations, fewer confident misses
Customer facing tutoring (concise answers)	Sonnet 4.6	Lower output token count, faster, cheaper at short outputs
Mixed domain enterprise reasoning	Either, route on domain	Domain routing beats picking one
Cost optimized batch reasoning	Fable 5	Cheaper at input heavy workloads, large context

The right answer for most enterprise deployments is not one model, it is a router. We have shipped production reasoning pipelines that route physics, physical chemistry, and long input workloads to Fable 5 and route biology, genetics, and short input conversational workloads to Sonnet 4.6. The router cost is negligible (a single small model classification call per request) and the cost savings against picking either model exclusively run 20 to 35 percent in our production deployments.

When This Applies to Your Stack

A scientific or regulatory pipeline that operates over long technical documents and emits structured analyses. Fable 5 wins on cost and on the physics and physical chemistry domains. Sonnet 4.6 wins on biology and genetics. A domain router captures both.

A customer facing reasoning workload (tutoring, explainer, troubleshooting) where users ask short questions and want concise focused answers. Sonnet 4.6 is the default. The latency and the output token cost advantage matter more than the 2.6 point aggregate gap.

A mixed domain enterprise AI deployment where the same model serves multiple internal teams with different domain needs. A simple classifier router beats picking one model. We ship this pattern in production and it captures real cost savings without the latency hit a more sophisticated routing layer would impose.

When This Does Not Apply

Workloads that need the absolute top of the reasoning leaderboard. Opus 4.8 and Gemini 3.5 Pro both sit above 78 percent on GPQA Diamond and are the right call when accuracy is the binding constraint and cost is not. Sonnet 4.6 and Fable 5 are the mid tier optimization, not the frontier ceiling.

Code generation workloads. Aider Polyglot and SWE-Bench Pro are the right benchmarks for that decision; GPQA Diamond is the wrong signal. Different model rankings apply.

Working with Contra Collective

We build production AI integration for teams shipping reasoning workloads at scale. If you are sizing a reasoning pipeline and want to know which model belongs at which step (or whether routing pays back the engineering cost), we run domain specific evaluations against your actual workload and produce a sized recommendation. Reach out via the Contra Collective contact page.

FAQ

Q: Is GPQA Diamond the right benchmark for picking a reasoning model? A: For graduate level science reasoning, yes. For general knowledge reasoning, MMLU-Pro is the broader signal. For coding reasoning, SWE-Bench Pro or Aider Polyglot is closer to the workload. Pick the benchmark that matches your task distribution; GPQA Diamond is the right pick when the workload is technical reasoning under uncertainty.

Q: Which model is better for production deployments overall? A: It depends on the workload mix and the cost structure. For input heavy workloads with long documents and structured outputs, Fable 5 wins on cost and is competitive on accuracy. For output heavy workloads with short questions and detailed answers, Sonnet 4.6 wins on cost and on latency. For mixed deployments, a domain router beats picking one model.

Q: How big is the latency gap in production? A: At 4K context with median output, Fable 5 is roughly 14 percent faster on time to first token, but it generates 2.9x more output tokens, so end to end response latency on long outputs is meaningfully slower. For latency sensitive customer facing workloads, Sonnet 4.6 with its more concise default style is the faster end to end pick.

Q: Does prompt engineering close the gap? A: Partially. Asking Sonnet 4.6 to externalize a chain of thought lifts its physics accuracy by roughly 4 points and closes the per domain gap. Asking Fable 5 to be concise drops its output token count by 30 to 50 percent at a small accuracy cost. The right answer is to pick the model whose default behavior matches the workload and not to fight the model with prompt engineering.

Q: How often do these benchmarks shift with new model releases? A: GPQA Diamond rankings shift roughly every two to three months as new model versions ship. We re run the eval suite on every new release and update the routing logic in production deployments. Treat any single point in time benchmark as a snapshot, not a permanent ranking.