Cookie preferences

Authentication and account storage support the service. Optional Google Analytics measures public pages only when you accept and analytics is enabled for this site. Private platform pages are excluded. You can change your choice here at any time.

Your analytics choice: not enabled without your choice.

Cookie details

ReasonBlocks

Evaluating ReasonBlocks
on WorkBench

ReasonBlocks builds specialized small models for specific tasks. We tested our proprietary training architecture on WorkBench, measuring accuracy, cost and latency on complex multi-step tasks with tool calling. All three ReasonBlocks models are on the Pareto frontier. Our flagship model, Tesseract, delivers Opus-level accuracy at 1/132 the cost.

WorkBench Model Performance

WorkBench accuracy by model and cost per attemptAccuracy versus cost per attempt on a logarithmic cost axis. The dashed line connects Pareto-efficient observations. Focus or tap a point for its accuracy and cost.707580859095100$0.0002$0.0005$0.001$0.003$0.01$0.03$0.1AccuracyClaude Haiku 4.5ReasonBlocks VertexGLM-5.3 FlashGPT-5.6 LunaGPT-5.6 SolGPT-5.6 TerraGLM-5.3ReasonBlocks PrismKimi K3Claude Sonnet 5Claude Opus 5ReasonBlocks TesseractClaude Fable 5.1

* Accuracy is measured on a randomly selected subset of 250 WorkBench tasks, with four trials per task. ReasonBlocks cost measurements assume maximum utilization of model capacity.

Cost / attempt

Trained on 390 tasks

ReasonBlocks Vertex: historical stock 29.2%, trained 84.4%. ReasonBlocks Prism: historical stock 52.4%, trained 93.8%. ReasonBlocks Tesseract: historical stock 64.9%, trained 96.6%.0255075100Accuracy29.2%0.0%ReasonBlocksVertex52.4%0.0%ReasonBlocksPrism64.9%0.0%ReasonBlocksTesseract

Our proprietary training method uses traces of task execution to improve how models work. We evaluate each version and use fresh traces to address its weaknesses, steadily improving performance over successive rounds of training.

WorkBench Model Latency

Median end-to-end WorkBench task latency by modelSeconds per task, lower is faster. One active task per model in a separate latency study. Errors are retained and timeouts are censored. Hover, focus, or tap a bar for its median latency.051015202530Median task latency (s)ReasonBlocksVertexReasonBlocksPrismReasonBlocksTesseractClaudeHaiku 4.5GPT-5.6LunaGPT-5.6TerraClaudeSonnet 5GPT-5.6SolClaudeOpus 5GLM-5.3FlashClaudeFable 5.1GLM-5.3Kimi K3

* Latency is measured with one active task at a time.

Our models are small and specialized, which helps them complete tasks quickly. The graph shows the median time to complete a task, including model responses and tool execution.

Footnotes

  1. We randomly selected 390 WorkBench tasks for training, 50 for development and 250 for evaluation. Accuracy is measured with four trials per evaluation task (1,000 trials total).
  2. ReasonBlocks costs reflect GPU rental at maximum utilization, excluding setup, idle time and other operating overhead.
  3. Latency is measured with one active task at a time.
  4. API models use high reasoning effort, except Haiku, which uses a 3,276-token thinking budget.

Keep the capability.
Cut the cost.

Request access Talk to ReasonBlocks