Claude Opus 5.5 vs Gemini 3.8 Flash
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Opus 5.561.2
Gemini 3.8 Flash60.9
GDPval-AA (work tasks) 2.1 GDPval-AA (work tasks) 2.1: Real work tasks from many professions (reports, spreadsheets, analyses), graded against expert work
Opus 5.554.2
Gemini 3.8 Flash46.1
AA Analyst Agent AA Analyst Agent: Data-analyst tasks: the model explores data with tools and reports findings
Opus 5.5 · max56.3
Gemini 3.8 Flash · high46.3
ITBench SRE ITBench SRE: IT incident tasks: diagnose and fix broken systems like a site-reliability engineer
Opus 5.5 · max38.2
Gemini 3.8 Flash · high52.5
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Opus 5.552.5
Gemini 3.8 Flash19.7
Terminal-Bench Science 0.1 Terminal-Bench Science 0.1: Agents complete expert-authored research workflows across five scientific domains in a terminal
Opus 5.543.3
Gemini 3.8 Flash · high10
AA-Briefcase 1.1 · Elo AA-Briefcase 1.1 · Elo: Agents produce professional deliverables; blind pairwise comparisons produce the source Elo rating
Opus 5.51,625.1
Gemini 3.8 Flash1,181.5
AA-LCR (long context) 1.1 AA-LCR (long context) 1.1: Questions that need reading and connecting facts across very long documents (around 100K tokens)
Opus 5.584.3
Gemini 3.8 Flash84
DeepSWE 1.1 DeepSWE 1.1: Agents change repository code to resolve software engineering tasks under the stated DeepSWE edition
Opus 5.5 · max72.3
Gemini 3.8 Flash71
FrontierCode 1.1 FrontierCode 1.1: Agents implement changes in real repositories, graded against maintainer-defined requirements
Opus 5.557.1
Gemini 3.8 Flash · high36.4
Harvey LAB-AA 1.1 Harvey LAB-AA 1.1: Legal agents work from case documents; every rubric criterion must pass and material hallucinations receive no credit
Opus 5.5 · max4.2
Gemini 3.8 Flash · high0
Furniture Assembly Furniture Assembly: Models inspect furniture assembly photos, identify incorrect builds and locate the mistaken step
Opus 5.5 · max83.3
Gemini 3.8 Flash · high31.7
τ³-Banking · AA τ³-Banking · AA: Agents retrieve banking knowledge and execute tools to resolve customer-support workflows
Opus 5.5—
Gemini 3.8 Flash45.8
SWE-bench Pro (v2 full) SWE-bench Pro (v2 full): Agents resolve repository issues under the network-locked v2 protocol; full and 51-task Hard results are separate
Opus 5.5—
Gemini 3.8 Flash · high94.9
SWE-bench Pro (v2 hard) SWE-bench Pro (v2 hard): Agents resolve repository issues under the network-locked v2 protocol; full and 51-task Hard results are separate
Opus 5.5—
Gemini 3.8 Flash · high58.8
EBR Bench EBR Bench: Agents learn to play Earthborne Rangers across repeated playthroughs
Opus 5.5 · max71.4
Gemini 3.8 Flash—
SpireClimb SpireClimb: Agents play Slay the Spire under the named game and difficulty setting
Opus 5.5 · max3.3
Gemini 3.8 Flash—
SpireClimb StS1 A20 Heart SpireClimb StS1 A20 Heart: Agents play Slay the Spire under the named game and difficulty setting
Opus 5.5 · max0
Gemini 3.8 Flash—
SpireClimb StS2 A10 SpireClimb StS2 A10: Agents play Slay the Spire under the named game and difficulty setting
Opus 5.5 · max6.7
Gemini 3.8 Flash—
| Benchmark | Opus 5.5 | Gemini 3.8 Flash |
|---|---|---|
| AutomationBench-AA | 61.2 | 60.9 |
| GDPval-AA (work tasks) 2.1 | 54.2 | 46.1 |
| AA Analyst Agent | 56.3 · max | 46.3 · high |
| ITBench SRE | 38.2 · max | 52.5 · high |
| Terminal-Bench 4.0 | 52.5 | 19.7 |
| Terminal-Bench Science 0.1 | 43.3 | 10 · high |
| AA-Briefcase 1.1 · Elo | 1,625.1 | 1,181.5 |
| AA-LCR (long context) 1.1 | 84.3 | 84 |
| DeepSWE 1.1 | 72.3 · max | 71 |
| FrontierCode 1.1 | 57.1 | 36.4 · high |
| Harvey LAB-AA 1.1 | 4.2 · max | 0 · high |
| Furniture Assembly | 83.3 · max | 31.7 · high |
| τ³-Banking · AA | — | 45.8 |
| SWE-bench Pro (v2 full) | — | 94.9 · high |
| SWE-bench Pro (v2 hard) | — | 58.8 · high |
| EBR Bench | 71.4 · max | — |
| SpireClimb | 3.3 · max | — |
| SpireClimb StS1 A20 Heart | 0 · max | — |
| SpireClimb StS2 A10 | 6.7 · max | — |
Specs
| Spec | Opus 5.5 | Gemini 3.8 Flash |
|---|---|---|
| Input, per 1M | $4 | !$0.75 |
| Output, per 1M | $20 | !$3.75 |
| Cost per task? | $1.34 | $0.93 |
| Speed, tok/s | 81 | — |
| Context | 1M | 1M |
| Weights | Closed | Closed |
| Released | Sep 2026 | Sep 2026 |
| See images | ✓ | ✓ |
| Read PDFs | ✓ | ✓ |
| Hear audio | — | ✓ |
| Watch video | — | ✓ |
| Call tools | ✓ | ✓ |
| Force a tool call | — | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ✓ | ✓ |
| Run code | ✓ | ✓ |
| Use a computer | ✓ | ? |
What breaks when you switch
Opus 5.5 → Gemini 3.8 Flash
Sampling parameters are ignored›
temperature, top_p and top_k are ignored now and will cause HTTP 400 in future generations; remove them from requests.
No minimal thinking level›
The minimal thinking level returns an error; avoid minimal.
Gemini 3.8 Flash → Opus 5.5
No forced tool use›
Forced tool_choice any or tool returns HTTP 400; remove the forced choice.
Thinking is always on›
Disabled thinking or a manual budget returns HTTP 400; omit thinking or use adaptive thinking.
Remove sampling parameters›
Non-default temperature, top_p or top_k returns HTTP 400; omit these parameters and use prompting.
No assistant prefill›
Assistant prefill returns HTTP 400; use structured outputs or system instructions, with system instructions on Bedrock.
Old computer tool rejected · Claude API / Google Cloud›
The computer_20251124 tool returns HTTP 400 on the Claude API and Google Cloud; use computer_toolset_20260801 there.
Return thinking blocks unmodified›
Edited, reordered or partially dropped thinking blocks return HTTP 400; echo assistant messages complete and unmodified with tool results.