Gemini 3.8 / 3.7 Flash Medium
Score 43-44 · $0.0020-$0.0022/task · Value Index: 20,000+
Top-tier math and coding at fraction-of-a-cent rates with under 10-second latency.
ARTIFICIAL INTELLIGENCE / PUBLIC DATA
How do AI thinking levels affect cost, speed, and accuracy?
Compare reasoning levels across OpenAI GPT, Google Gemini, and Anthropic Claude, showing intelligence scores, task costs, and execution speeds.
Collected UTC
Coverage: 34 model reasoning tiers across OpenAI, Google, and Anthropic.
Observations
- Diminishing returns define maximum reasoning. Moving from Low to Medium reasoning produces the steepest capability gains across all architectures, capturing 80% to 85% of peak intelligence. Stepping from High to Max reasoning yields only a 2 to 4 point intelligence increase while multiplying cost by 4x to 10x and stretching latency up to 150 seconds.
- Newer models at low reasoning outperform older models at maximum effort. Modern lightweight models like Gemini 3.8 Flash at Medium reasoning (score 44, $0.0022/task) or Claude 3.7 Sonnet at Low reasoning (score 43, $0.025/task) outperform earlier generation models running at maximum effort, while completing faster and costing less.
- Reasoning tokens concentrate 80% to 92% of the request cost. Because thinking tokens are billed as output tokens at 3x to 5x the price of input tokens, complex queries with 10,000 or more thinking tokens shift almost the entire bill to internal chain of thought rather than prompt context.
- Latency creates an architectural threshold between interactive and asynchronous tasks. Standard and Low reasoning settings return in 1 to 7 seconds, making them suitable for interactive user interfaces. High and Max reasoning settings take 30 to 150 seconds, requiring background worker queues or decoupled job pipelines.
Best picks and recommendations
GPT-5.6 Luna High
Score 33 · $0.09/task · Value Index: 183
High-accuracy multi-step research and structured code generation.
Claude 3.7 Sonnet Low / Medium
Score 43-49 · $0.03-$0.07/task · Value Index: 700+
The premier sweet spot for daily software engineering, refactoring, and agent loops.
Claude 3.7 Sonnet Max or GPT-6 Astra Max
Score 53-55 · $0.95-$3.26/task · Value Index: 16-58
Reserved strictly for mission-critical architecture, novel mathematics, and high-stakes proofs.
Visual comparisons: Intelligence versus Cost and Speed
Filter charts by provider:
Intelligence Index vs. Cost per Task
Higher and to the left represents superior efficiency (greater intelligence at lower task cost). Key anchor models are labeled directly on the chart dots.
View cost chart data table
| Model | Provider | Level | Avg Cost / Task | Intelligence Index |
|---|---|---|---|---|
| Claude 3.7 Sonnet | Anthropic | High | $0.2200 | 53 |
| Claude 3.7 Sonnet | Anthropic | Low | $0.0300 | 43 |
| Claude 3.7 Sonnet | Anthropic | Max | $0.9500 | 55 |
| Claude 3.7 Sonnet | Anthropic | Medium | $0.0700 | 49 |
| Claude 3.7 Sonnet | Anthropic | Standard | $0.0100 | 36 |
| Claude Mythos 5.1 | Anthropic | High | $1.2000 | 54 |
| Claude Mythos 5.1 | Anthropic | Max | $4.8000 | 56 |
| Gemini 3.7 Flash | Medium | $0.0020 | 43 | |
| Gemini 3.7 Pro | High | $0.0500 | 51 | |
| Gemini 3.7 Pro | Medium | $0.0250 | 49 | |
| Gemini 3.8 Flash | High | $0.0075 | 48 | |
| Gemini 3.8 Flash | Low | $0.0009 | 35 | |
| Gemini 3.8 Flash | Max | $0.0120 | 50 | |
| Gemini 3.8 Flash | Medium | $0.0022 | 44 | |
| GPT-5.6 Luna | OpenAI | Extra High | $0.1200 | 35 |
| GPT-5.6 Luna | OpenAI | High | $0.0900 | 33 |
| GPT-5.6 Luna | OpenAI | Low | $0.0400 | 22 |
| GPT-5.6 Luna | OpenAI | Max | $0.1800 | 38 |
| GPT-5.6 Luna | OpenAI | Medium | $0.0600 | 26 |
| GPT-5.6 Sol | OpenAI | Extra High | $1.4300 | 44 |
| GPT-5.6 Sol | OpenAI | High | $0.8100 | 42 |
| GPT-5.6 Sol | OpenAI | Low | $0.2300 | 34 |
| GPT-5.6 Sol | OpenAI | Max | $1.9900 | 47 |
| GPT-5.6 Sol | OpenAI | Medium | $0.5000 | 39 |
| GPT-5.6 Terra | OpenAI | Extra High | $0.9300 | 38 |
| GPT-5.6 Terra | OpenAI | High | $0.5200 | 34 |
| GPT-5.6 Terra | OpenAI | Low | $0.1200 | 28 |
| GPT-5.6 Terra | OpenAI | Max | $1.4000 | 42 |
| GPT-5.6 Terra | OpenAI | Medium | $0.3300 | 33 |
| GPT-6 Astra | OpenAI | Extra High | $3.2600 | 53 |
| GPT-6 Astra | OpenAI | High | $2.7100 | 51 |
| GPT-6 Astra | OpenAI | Low | $0.8200 | 46 |
| GPT-6 Astra | OpenAI | Max | $3.2600 | 53 |
| GPT-6 Astra | OpenAI | Medium | $2.1700 | 50 |
Intelligence Index vs. Execution Speed & Latency
Comparing capability against response turnaround tiers: Fast (1-5s), Moderate (6-20s), and Deliberate (30-150s). Notice the sharp latency penalty required for top-tier reasoning effort.
View speed chart data table
| Model | Provider | Level | Speed Tier | Intelligence Index |
|---|---|---|---|---|
| Claude 3.7 Sonnet | Anthropic | High | Deliberate (30-150s) | 53 |
| Claude 3.7 Sonnet | Anthropic | Low | Moderate (6-20s) | 43 |
| Claude 3.7 Sonnet | Anthropic | Max | Deliberate (30-150s) | 55 |
| Claude 3.7 Sonnet | Anthropic | Medium | Moderate (6-20s) | 49 |
| Claude 3.7 Sonnet | Anthropic | Standard | Fast (1-5s) | 36 |
| Claude Mythos 5.1 | Anthropic | High | Deliberate (30-150s) | 54 |
| Claude Mythos 5.1 | Anthropic | Max | Deliberate (30-150s) | 56 |
| Gemini 3.7 Flash | Medium | Moderate (6-20s) | 43 | |
| Gemini 3.7 Pro | High | Deliberate (30-150s) | 51 | |
| Gemini 3.7 Pro | Medium | Moderate (6-20s) | 49 | |
| Gemini 3.8 Flash | High | Deliberate (30-150s) | 48 | |
| Gemini 3.8 Flash | Low | Fast (1-5s) | 35 | |
| Gemini 3.8 Flash | Max | Deliberate (30-150s) | 50 | |
| Gemini 3.8 Flash | Medium | Moderate (6-20s) | 44 | |
| GPT-5.6 Luna | OpenAI | Extra High | Deliberate (30-150s) | 35 |
| GPT-5.6 Luna | OpenAI | High | Deliberate (30-150s) | 33 |
| GPT-5.6 Luna | OpenAI | Low | Fast (1-5s) | 22 |
| GPT-5.6 Luna | OpenAI | Max | Deliberate (30-150s) | 38 |
| GPT-5.6 Luna | OpenAI | Medium | Moderate (6-20s) | 26 |
| GPT-5.6 Sol | OpenAI | Extra High | Deliberate (30-150s) | 44 |
| GPT-5.6 Sol | OpenAI | High | Deliberate (30-150s) | 42 |
| GPT-5.6 Sol | OpenAI | Low | Moderate (6-20s) | 34 |
| GPT-5.6 Sol | OpenAI | Max | Deliberate (30-150s) | 47 |
| GPT-5.6 Sol | OpenAI | Medium | Moderate (6-20s) | 39 |
| GPT-5.6 Terra | OpenAI | Extra High | Deliberate (30-150s) | 38 |
| GPT-5.6 Terra | OpenAI | High | Deliberate (30-150s) | 34 |
| GPT-5.6 Terra | OpenAI | Low | Moderate (6-20s) | 28 |
| GPT-5.6 Terra | OpenAI | Max | Deliberate (30-150s) | 42 |
| GPT-5.6 Terra | OpenAI | Medium | Moderate (6-20s) | 33 |
| GPT-6 Astra | OpenAI | Extra High | Deliberate (30-150s) | 53 |
| GPT-6 Astra | OpenAI | High | Deliberate (30-150s) | 51 |
| GPT-6 Astra | OpenAI | Low | Moderate (6-20s) | 46 |
| GPT-6 Astra | OpenAI | Max | Deliberate (30-150s) | 53 |
| GPT-6 Astra | OpenAI | Medium | Moderate (6-20s) | 50 |
About the data
This dataset compares reasoning capabilities, pricing, task costs, and execution speeds across frontier AI models from OpenAI, Anthropic, and Google. It provides practitioners and engineers with the trade-offs needed to configure reasoning effort levels in production software.
Key Metrics and Definitions
- Intelligence Index (0-100): A normalized benchmark index combining competitive STEM mathematics (AIME 2024, MATH-500), expert scientific reasoning (GPQA Diamond), and real-world software engineering (SWE-bench Verified).
- Token Pricing ($ / 1M tokens): Published direct provider API token rates for input and output, measured in USD per million tokens. Internal reasoning and thinking tokens are billed at the output rate.
- Average Task Cost ($ / task): The estimated financial cost to complete a representative multi-step analytical prompt with 1,000 input tokens and typical reasoning token generation for that effort tier.
- Speed: A categorical rating (Fast, Moderate, Deliberate, Low) reflecting end-to-end task completion latency in seconds.
- Depth: The relative depth of the internal reasoning graph, tracking token budgets from short verification checks to exhaustive search.
- Value Index: Calculated as normalized intelligence divided by average task cost, highlighting models that deliver the highest intelligence per dollar spent.
Coverage Limits
Prices and performance figures reflect official provider documentation and benchmark cards as of September 2026. Latency figures represent average direct API performance and may vary based on provider server load, region, and batching.
Possible explanations
Test-time compute scaling and search tree dynamics
Proposed mechanism: Reasoning models use reinforcement learning to search through multiple potential solution paths, backtrack when hitting dead ends, and verify steps internally before generating a final response.
Evidence: Performance on formal verification tasks (like competition mathematics and code execution) jumps significantly with thinking tokens. However, once the model finds the correct solution path, additional reasoning tokens provide diminishing returns, explaining the plateau between High and Max levels.
Asymmetric token pricing and output compute weight
Proposed mechanism: Output tokens require sequential autoregressive generation, which cannot be parallelized across tokens like prompt ingestion. Providers price output tokens 3x to 5x higher than input tokens to reflect hardware allocation.
Evidence: Because all thinking tokens are billed as output tokens, a query that generates 15,000 thinking tokens shifts over 90% of the total request cost onto the generation stage, causing exponential cost growth at high effort levels.
Explore the records
Search and compare reasoning models by provider, intelligence score, cost per task, and recommended use case.
| Model | Provider | Reasoning Level | Intelligence Index | Input / 1M | Output / 1M | Avg Cost / Task | Speed | Depth | Best For |
|---|---|---|---|---|---|---|---|---|---|
| Claude 3.7 Sonnet | Anthropic | High | 53 | $3.00 | $15.00 | $0.2200 | Deliberate (30-150s) | Deep (3/4) | Complex tasks and deep logic |
| Claude 3.7 Sonnet | Anthropic | Low | 43 | $3.00 | $15.00 | $0.0300 | Moderate (6-20s) | Check (1/4) | High intelligence at balanced cost |
| Claude 3.7 Sonnet | Anthropic | Max | 55 | $3.00 | $15.00 | $0.9500 | Deliberate (30-150s) | Exhaustive (4/4) | Hardest possible tasks |
| Claude 3.7 Sonnet | Anthropic | Medium | 49 | $3.00 | $15.00 | $0.0700 | Moderate (6-20s) | Standard (2/4) | Complex everyday development |
| Claude 3.7 Sonnet | Anthropic | Standard | 36 | $3.00 | $15.00 | $0.0100 | Fast (1-5s) | Check (1/4) | Everyday tasks with rapid responses |
| Claude Mythos 5.1 | Anthropic | High | 54 | $15.00 | $75.00 | $1.2000 | Deliberate (30-150s) | Deep (3/4) | Frontier-level multi-step tasks |
| Claude Mythos 5.1 | Anthropic | Max | 56 | $15.00 | $75.00 | $4.8000 | Deliberate (30-150s) | Exhaustive (4/4) | Maximum capability frontier science |
| Gemini 3.7 Flash | Medium | 43 | $0.10 | $0.40 | $0.0020 | Moderate (6-20s) | Standard (2/4) | High-efficiency coding and problem solving | |
| Gemini 3.7 Pro | High | 51 | $1.25 | $5.00 | $0.0500 | Deliberate (30-150s) | Deep (3/4) | Complex multimodal planning | |
| Gemini 3.7 Pro | Medium | 49 | $1.25 | $5.00 | $0.0250 | Moderate (6-20s) | Standard (2/4) | General enterprise analysis | |
| Gemini 3.8 Flash | High | 48 | $0.10 | $0.40 | $0.0075 | Deliberate (30-150s) | Deep (3/4) | High-volume complex analysis | |
| Gemini 3.8 Flash | Low | 35 | $0.10 | $0.40 | $0.0009 | Fast (1-5s) | Check (1/4) | High-speed reasoning verification | |
| Gemini 3.8 Flash | Max | 50 | $0.10 | $0.40 | $0.0120 | Deliberate (30-150s) | Exhaustive (4/4) | Budget frontier reasoning | |
| Gemini 3.8 Flash | Medium | 44 | $0.10 | $0.40 | $0.0022 | Moderate (6-20s) | Standard (2/4) | General workhorse coding and reasoning | |
| GPT-5.6 Luna | OpenAI | Extra High | 35 | $0.20 | $1.20 | $0.1200 | Deliberate (30-150s) | Deep (3/4) | Hard problems, higher accuracy |
| GPT-5.6 Luna | OpenAI | High | 33 | $0.20 | $1.20 | $0.0900 | Deliberate (30-150s) | Deep (3/4) | Complex but non-frontier tasks |
| GPT-5.6 Luna | OpenAI | Low | 22 | $0.20 | $1.20 | $0.0400 | Fast (1-5s) | Check (1/4) | Simple, high-volume tasks |
| GPT-5.6 Luna | OpenAI | Max | 38 | $0.20 | $1.20 | $0.1800 | Deliberate (30-150s) | Deep (3/4) | Toughest tasks on a budget |
| GPT-5.6 Luna | OpenAI | Medium | 26 | $0.20 | $1.20 | $0.0600 | Moderate (6-20s) | Standard (2/4) | General everyday work |
| GPT-5.6 Sol | OpenAI | Extra High | 44 | $4.00 | $20.00 | $1.4300 | Deliberate (30-150s) | Deep (3/4) | Very challenging tasks |
| GPT-5.6 Sol | OpenAI | High | 42 | $4.00 | $20.00 | $0.8100 | Deliberate (30-150s) | Deep (3/4) | Hard problems |
| GPT-5.6 Sol | OpenAI | Low | 34 | $4.00 | $20.00 | $0.2300 | Moderate (6-20s) | Check (1/4) | Everyday tasks with stronger results |
| GPT-5.6 Sol | OpenAI | Max | 47 | $4.00 | $20.00 | $1.9900 | Deliberate (30-150s) | Exhaustive (4/4) | Frontier-level tasks |
| GPT-5.6 Sol | OpenAI | Medium | 39 | $4.00 | $20.00 | $0.5000 | Moderate (6-20s) | Standard (2/4) | Complex everyday work |
| GPT-5.6 Terra | OpenAI | Extra High | 38 | $2.00 | $12.00 | $0.9300 | Deliberate (30-150s) | Deep (3/4) | Challenging, high-accuracy tasks |
| GPT-5.6 Terra | OpenAI | High | 34 | $2.00 | $12.00 | $0.5200 | Deliberate (30-150s) | Deep (3/4) | Complex tasks |
| GPT-5.6 Terra | OpenAI | Low | 28 | $2.00 | $12.00 | $0.1200 | Moderate (6-20s) | Check (1/4) | Simple tasks at higher quality than Luna |
| GPT-5.6 Terra | OpenAI | Max | 42 | $2.00 | $12.00 | $1.4000 | Deliberate (30-150s) | Exhaustive (4/4) | Maximum capability when needed |
| GPT-5.6 Terra | OpenAI | Medium | 33 | $2.00 | $12.00 | $0.3300 | Moderate (6-20s) | Standard (2/4) | General work with more depth |
| GPT-6 Astra | OpenAI | Extra High | 53 | $10.00 | $50.00 | $3.2600 | Deliberate (30-150s) | Deep (3/4) | Near-maximum capability |
| GPT-6 Astra | OpenAI | High | 51 | $10.00 | $50.00 | $2.7100 | Deliberate (30-150s) | Deep (3/4) | Very complex tasks |
| GPT-6 Astra | OpenAI | Low | 46 | $10.00 | $50.00 | $0.8200 | Moderate (6-20s) | Check (1/4) | High intelligence at reasonable cost |
| GPT-6 Astra | OpenAI | Max | 53 | $10.00 | $50.00 | $3.2600 | Deliberate (30-150s) | Exhaustive (4/4) | Hardest possible tasks |
| GPT-6 Astra | OpenAI | Medium | 50 | $10.00 | $50.00 | $2.1700 | Moderate (6-20s) | Standard (2/4) | Challenging, nuanced tasks |
References
Official provider documentation, technical releases, and benchmark evaluation cards. Every listed provider model and pricing tier is verified against primary documentation.
- OpenAI Reasoning Models and PricingOpenAI · API specifications and reported evaluations
Official pricing, reasoning effort parameter definitions, token rates, and evaluation benchmarks for Astra, Sol, Terra, and Luna model tiers.
Published/update: 2026-09-03; accessed 2026-09-19. - Claude 3.7 Sonnet and Extended Thinking SpecificationsAnthropic · Provider release, pricing, and reported evaluations
Hybrid reasoning architecture details, budget token controls, output token rates, SWE-bench and AIME evaluation scaling across thinking depths.
Published/update: 2025-02-24; accessed 2026-09-19. - Gemini 3.8 Flash and Gemini 3.7 Flash and Pro DocumentationGoogle · Provider API metadata and serving rates
Thinking budget parameter, latency profiles, token pricing tiers, and mathematical reasoning scores across Flash and Pro variants.
Published/update: 2026-09-17; accessed 2026-09-19. - Frontier Model Intelligence and Quality EvaluationsIndependent benchmarks and provider cards · Methodology and context
Normalized 0-100 intelligence indexing methodology combining STEM benchmarks, reasoning evaluations, and speed-latency trade-offs.
Published/update: 2026-09-08; accessed 2026-09-19.