Across 27 comparable benchmarks, Gemini 3.8 Flash has better reported results on 20 and worse on 7. 1 further shared result is listed below, each with the reason it is not counted.
Strong evidence27 of 28 shared results countedHow we compare
Price basis: Current effective rates. The price section below shows the applicable periods.
| Fact | Mistral Large 4 | Gemini 3.8 Flash |
|---|---|---|
| Context window | 1M tokens | 1M tokens |
| Input / 1M | $0.700 | $0.750 |
| Output / 1M | $2.10 | $3.75 |
| Blended (4:1) | $0.526 | $0.432 |
| Status | Preview | Available |
| Released | October 6, 2026 | September 2, 2026 |
Blended is 4:1 input to output with 90% cache hits at listed rates — set your own mix.
Mean of 19 shared benchmarks grouped by capability on a 0–100 scale. How averages are built
Gemini 3.8 Flash ahead on 6 of 10 counted
Gemini 3.8 Flash ahead on 2 of 3 counted
Gemini 3.8 Flash ahead on 3 of 4 counted
Gemini 3.8 Flash ahead on 1 of 1 counted
Gemini 3.8 Flash ahead on 4 of 5 counted
No shared comparable results for Abstract reasoning, Computer & tool use, Research & search, Long context, Cybersecurity.
Published scores, with matching modes where available. Gaps show Mistral Large 4 relative to Gemini 3.8 Flash. Reading the results
| Benchmark | Mistral Large 4 | Gemini 3.8 Flash | Gap |
|---|---|---|---|
| AA Cyber IndexRepository Engineering · Composite index (0–100) | 49.5Thinking · tools | 25.1≠Thinking High · tools | 1.97×#7 vs #17 of 22 |
| Code MigrationMaintenance & Optimization · Hidden-test pass rate (%) | 30.6Thinking High · tools | 36.5Thinking High · tools | −6.0 pts#43 vs #37 of 79 |
| DeepSWERepository Engineering · Pass@1 (DeepSWE v1.1) | 61.7Thinking · tools | 73.8≠Thinking High · tools | −12.1 pts#45 vs #14 of 73 |
| FrontierCode 1.1Code Generation & Editing · Weighted rubric score (%)Not counted · Different or unrecorded test environments — excluded from the tally | 39.4Thinking High · tools· frontiercode-1-1-main | 56.3≠Unspecified Reasoning · tools· frontiercode-1-1-extended | —#30 vs #10 of 51 |
| IOI (Vals v2)Algorithmic Coding · Subtask points (%) | 45.3Thinking High · tools | 56.9Thinking High · tools | −11.7 pts#32 vs #20 of 45 |
| SWE-Atlas-QnARepository Engineering · Task Resolve Rate (%) | 59.4Thinking · tools | 47.0Thinking · tools | +12.4 pts#4 vs #11 of 17 |
| Terminal-Bench 4.0Agentic Development · Resolution rate (%) | 28.3Thinking · tools | 19.7≠Thinking Medium · tools | +8.6 pts#40 vs #49 of 90 |
| Terminal-Bench 4.1Agentic Development · Resolution rate (%) | 27.6Max · tools | 18.2≠Thinking High · tools | +9.4 pts#26 vs #33 of 48 |
| Vibe Code Bench 1-100Agentic Development · Accuracy (%) | 14.3Thinking High · tools | 18.8Thinking High · tools | −4.5 pts#20 vs #13 of 25 |
| Vibe Code Bench v1.1Code Generation & Editing · Pass rate (%) | 78.4Thinking High · tools | 78.7Thinking High · tools | −0.3 pts#36 vs #35 of 121 |
| WeirdML v3ML Engineering · Official score (0-1; higher is better) | 0.044Thinking High · tools | 0.074Thinking High · tools | 0.60×#25 vs #19 of 26 |
| FrontierMath Tier 4 v2Mathematics · Accuracy (verification_code) | 26.8Thinking High · tools | 22.0Thinking High · tools | +4.9 pts#40 vs #49 of 73 |
| FrontierMath v2Mathematics · Accuracy (verification_code) | 58.6Thinking High · tools | 68.4Thinking High · tools | −9.8 pts#46 vs #33 of 85 |
| ProofBench v1.1 (Lean 4)Mathematics · Accuracy (%) | 10.0Thinking High · tools | 48.0Thinking High · tools | −38.0 pts#49 vs #35 of 54 |
| MysteryMechanismScientific Reasoning · Accuracy (%) | 12.2Thinking High · tools | 36.5Thinking High · tools | −24.3 pts#32 vs #11 of 32 |
| BioMysteryBenchBiology & Genomics · Accuracy (%) | 67.0Thinking High · tools | 62.2Thinking High · tools | +4.8 pts#21 vs #24 of 28 |
| MedCodeClinical Workflows · Accuracy (%) | 40.7Thinking High · no tools | 48.1≠Thinking High · tools | −7.4 pts#58 vs #30 of 97 |
| MedScribeClinical Workflows · Rubric pass rate (%) | 80.4Thinking High · no tools | 84.5≠Thinking High · tools | −4.1 pts#53 vs #34 of 99 |
| MMMU-ProVisual Understanding · Accuracy | 76.4Thinking · no tools | 85.6≠Thinking High · no tools | −9.2 pts#55 vs #7 of 136 |
| EMB (Excel Modeling)Finance · Accuracy (%) | 56.2Thinking High · tools | 72.2Thinking High · tools | −16.0 pts#49 vs #10 of 78 |
| Finance Agent v2Finance · Score | 54.7Thinking High · tools | 61.4≠Thinking · tools | 0.89×#24 vs #2 of 75 |
| Harvey's Legal Agent BenchmarkLegal · Accuracy (%) | 15.8Thinking High · tools | 10.0Thinking High · tools | +5.8 pts#6 vs #13 of 52 |
| Legal Research BenchLegal · Score | 31.7Thinking High · tools | 38.9Thinking High · tools | 0.81×#52 vs #39 of 91 |
| Tax Agent BenchFinance · Accuracy (%, rubric × citation quality) | 63.3Thinking High · tools | 66.8Thinking High · tools | −3.5 pts#14 vs #9 of 74 |
| AA Intelligence Index v5.0.1Capability Indices · Index (0-100) | 25.3Thinking · tools | 30.8≠Thinking High · tools | 0.82×#39 vs #29 of 62 |
| LiveBenchCross-capability Suites · Accuracy | 71.8Thinking High · no tools | 75.8Thinking High · no tools | −4.0 pts#53 vs #30 of 106 |
| LiveBench 2.0Cross-capability Suites · Overall score (equal-weight mean of 7 suite scores, 0-100) | 47.7Thinking High · tools | 67.3Thinking High · tools | 0.71×#47 vs #29 of 51 |
| Vals Index v2.1Capability Indices · GDP-weighted accuracy (%) | 48.0Thinking High · tools | 54.8Thinking High · tools | −6.8 pts#30 vs #15 of 38 |
Repository Engineering · Composite index (0–100)
Maintenance & Optimization · Hidden-test pass rate (%)
Repository Engineering · Pass@1 (DeepSWE v1.1)
Code Generation & Editing · Weighted rubric score (%)
Not counted · Different or unrecorded test environments — excluded from the tally
Algorithmic Coding · Subtask points (%)
Repository Engineering · Task Resolve Rate (%)
Agentic Development · Resolution rate (%)
Agentic Development · Resolution rate (%)
Agentic Development · Accuracy (%)
Code Generation & Editing · Pass rate (%)
ML Engineering · Official score (0-1; higher is better)
Mathematics · Accuracy (verification_code)
Mathematics · Accuracy (verification_code)
Mathematics · Accuracy (%)
Scientific Reasoning · Accuracy (%)
Biology & Genomics · Accuracy (%)
Clinical Workflows · Accuracy (%)
Clinical Workflows · Rubric pass rate (%)
Visual Understanding · Accuracy
Finance · Accuracy (%)
Finance · Score
Legal · Accuracy (%)
Legal · Score
Finance · Accuracy (%, rubric × citation quality)
Capability Indices · Index (0-100)
Cross-capability Suites · Accuracy
Cross-capability Suites · Overall score (equal-weight mean of 7 suite scores, 0-100)
Capability Indices · GDP-weighted accuracy (%)
A dash means no result is available.
| Benchmark | Mistral Large 4 | Gemini 3.8 Flash |
|---|---|---|
| CritPt | — | 18.3 |
| FrontierMath Erdős | — | 0.0 |
| GPQA Diamond | — | 95.3 |
| HLE | — | 44.5 |
| HLE-Verified | — | 54.9 |
| Riemann-bench | — | 51.2 |
| SciCode | — | 56.6 |
| Terminal-Bench-Science 0.1 | — | 12.4 |
| APEX-SWE | — | 36.3 |
| CursorBench 4.0 | — | 39.6 |
| FrontierCode 2.0 | 8.8 | — |
| FrontierSWE v2 | — | 19.6 |
| SWE-Bench Pro V2 | — | 94.9 |
| SWE-Bench Pro V2 Hard | — | 58.8 |
| Terminal-Bench 2.1 | — | 89.4 |
| APEX-Agents | — | 64.3 |
| OSWorld 2.0 | — | 59.0 |
| Agent Arena | — | 2.3 |
| Agent Arena Index | — | 1118 |
| ECI | — | 157 |
| Vals Index | — | 62.3 |
| Vals Index v2.2 | — | 56.0 |
| Creative Writing | — | 1744 |
| Hypercritica Model Bench | — | 72.1 |
| AA-LCR | — | 84.0 |
| CharXiv RQ | — | 86.2 |
| GDP.pdf | — | 22.8 |
| GeoBench ACW | — | 86.0 |
| LVBench | — | 87.8 |
| AA-AnalystAgent | — | 46.3 |
| AutomationBench | — | 29.7 |
| Public Benefits Bench v1.1 | — | 65.3 |
| SAGE | — | 35.1 |
| τ³-Banking | — | 45.8 |
| ARC-AGI-1 | — | 98.5 |
| ARC-AGI-2 | — | 89.2 |
| ARC-AGI-3 | — | 35.0 |
| GoBench | — | 1616 |
| SimpleBench | — | 82.4 |
| Vending-Bench 2 | — | 5094 |
| BioMysteryBench (Human-difficult) | — | 56.5 |
| BioMysteryBench (Human-solvable) | — | 88.8 |
| LABBench2 | — | 86.2 |
Both models publish peak and off-peak prices. Each period is compared separately using the same token mix; hours follow each provider’s schedule.
Drag the handle above the bar for cache hits, the one below it for the input : output mix · 5% steps
Blended price = effective input rate at 90% cache hits × 80% + output rate × 20%
Mistral Large 4 costs39%lessthan Gemini 3.8 Flash per million tokens.Using current effective rates at this token mix.
Mistral Large 4: Standard pricing. Mistral official pricing (docs.mistral.ai/inference/pricing) verified 2026-11-13; price reduced from $1.36 Mistral official pricing (docs.mistral.ai/inference/pricing) verified 2026-11-13; price reduced from $4.18
Gemini 3.8 Flash: Listed text rates; pricing tier not confirmed.
| Price / 1M tokens | Mistral Large 4 | Gemini 3.8 Flash | Ratio |
|---|---|---|---|
| Input | $0.70 | $0.75 | 0.93× |
| Output | $2.10 | $3.75 | 0.56× |
| Cached input | $0.07 | $0.075 | 0.93× |
| Cache write | — | — | — |
| Blended | $0.526 | $0.864 | 0.61× |
Mistral Large 4 costs70%lessthan Gemini 3.8 Flash per million tokens — 3.3× cheaper.Using standard list rates after promotions end at this token mix.
Mistral Large 4: Standard pricing. Mistral official pricing (docs.mistral.ai/inference/pricing) verified 2026-11-13; price reduced from $1.36 Mistral official pricing (docs.mistral.ai/inference/pricing) verified 2026-11-13; price reduced from $4.18
Gemini 3.8 Flash: Standard pricing. Google Gemini API official pricing (ai.google.dev/gemini-api/docs/pricing) verified 2026-10-21; paid tier; standard tier; scheduled price starting 2027-01-01
| Price / 1M tokens | Mistral Large 4 | Gemini 3.8 Flash | Ratio |
|---|---|---|---|
| Input | $0.70 | $1.50 | 0.47× |
| Output | $2.10 | $7.50 | 0.28× |
| Cached input | $0.07 | $0.15 | 0.47× |
| Cache write | — | — | — |
| Blended | $0.526 | $1.73 | 0.30× |
Estimates use the selected token mix. Cache writes, tool fees and long-context surcharges are excluded.
Reported cost per task, paired with the score from the same run. Settings and sources may differ; costs do not affect the verdict.
Mistral Large 4 relative to Gemini 3.8 Flash
Summary: 6 of 7 shared benchmarks have matching recorded settings and sources. The average is a geometric mean of 6 positive-cost ratios, one equal weight per benchmark. This benchmark set can include related tests and indexes; it is not a workload cost estimate.
Mistral Large 4 7.3× more expensive
Mistral Large 4 4.4× cheaper
Mistral Large 4 3.2× cheaper
Mistral Large 4 2.8× cheaper
Mistral Large 4 2.4× more expensive
Mistral Large 4 1.2× cheaper
Settings or sources differ
Mistral Large 4 7.3× more expensive
Mistral Large 4 4.4× cheaper
Mistral Large 4 3.2× cheaper
Mistral Large 4 2.8× cheaper
Mistral Large 4 2.4× more expensive
Mistral Large 4 1.2× cheaper
Settings or sources differ
Effort, protocol, aggregation and source match; other run conditions may differ.
Mistral Large 4
Gemini 3.8 Flash
Effort, protocol, aggregation and source match; other run conditions may differ.
Mistral Large 4
Gemini 3.8 Flash
Effort, protocol, aggregation and source match; other run conditions may differ.
Mistral Large 4
Gemini 3.8 Flash
Settings or sources differ or are not recorded; interpret the comparison with care.
Mistral Large 4
Gemini 3.8 Flash
Effort, protocol, aggregation and source match; other run conditions may differ.
Mistral Large 4
Gemini 3.8 Flash
Effort, protocol, aggregation and source match; other run conditions may differ.
Mistral Large 4
Gemini 3.8 Flash
Effort, protocol, aggregation and source match; other run conditions may differ.
Mistral Large 4
Gemini 3.8 Flash
| Benchmark | Mistral Large 4 | Gemini 3.8 Flash |
|---|---|---|
| AA-LCR | Not reported | |
| Agent Arena | Not reported | $0.7652 / task Median · Net improvement index (percentage points vs. average model): 2.3 Thinking High (With Tools + Internet) |
| Agent Arena Index | Not reported | $0.7486 / task Median · Agent Arena Index rating (index points): 1118 Thinking High (With Tools + Internet) |
| ARC-AGI-1 | Not reported | |
| ARC-AGI-2 | Not reported | $0.2478 / task Average · ARC-AGI-2 Score: 82.9 Thinking Medium (No Tools) |
| AutomationBench | Not reported | $0.62 / task Average · Pass Rate: 29.7 Thinking High (With Tools) · zapier-private-heldout $0.55 / task Average · Pass Rate: 29.7 Thinking Medium (With Tools) · zapier-private-heldout |
| BioMysteryBench | Not reported | |
| Code Migration | Not reported | |
| CursorBench 4.0 | Not reported | $4.70 / task Average · Task score: 39.6 Thinking High (With Tools) · cursorbench-4-0-cursor-agent $4.06 / task Average · Task score: 37.3 Thinking Medium (With Tools) · cursorbench-4-0-cursor-agent |
| DeepSWE | Not reported | |
| EMB (Excel Modeling) | Not reported | |
| Finance Agent v2 | Not reported | |
| FrontierCode 2.0 | Not reported | |
| FrontierSWE v2 | Not reported | |
| Harvey's Legal Agent Benchmark | Not reported | |
| IOI (Vals v2) | Not reported | |
| Legal Research Bench | Not reported | |
| MedCode | Not reported | |
| MedScribe | Not reported | |
| MysteryMechanism | Not reported | |
| ProofBench v1.1 (Lean 4) | Not reported | |
| Tax Agent Bench | $2.7212 / task Average · Accuracy (%, rubric × citation quality): 63.3 Thinking High (With Tools) | Not reported |
| Vals Index v2.2 | Not reported | |
| Vibe Code Bench v1.1 | Not reported |
| Attribute | Mistral Large 4 | Gemini 3.8 Flash |
|---|---|---|
| Availability | Preview | Available |
| Provider | MistralAI | Google Deep Mind |
| Model type | Multimodal model | Multimodal model |
| First release | October 6, 2026 | September 2, 2026 |
| Context window | 1M tokens | 1M tokens |
| Max output | — | 64K tokens |
| Total parameters | 1.1T | — |
| Active parameters | 49B | — |
| Architecture | Mixture of Experts (MoE) | Undisclosed |
| Documented capabilities | Reasoning, Multilingual, Function calling, Tool use, Structured output, Long context, Coding | Reasoning, Coding, Function calling, Tool use, Structured output, Computer use, Web search, Long context |
| Input modalities | Text, Image | Text, Image, Audio, Video |
| Output modalities | Text | Text |
| Weights | Closed / API only | Closed / API only |
| Commercial use | Proprietary | Proprietary |
These are reported-score comparisons, not controlled head-to-head tests.
What people found comparing these models, or using either one. Comments also appear on the models' own pages.
Used Mistral Large 4 or Gemini 3.8 Flash? Sign in with Google to share what worked, what did not, and which API or setup you tested.
No community comments on either model yet.
Alternatives with the most benchmark overlap for this matchup.