Claude Sonnet 5.5 vs Gemini 4 Argon
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Sonnet 5.559.4
Gemini 4 Argon77.5
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
Sonnet 5.552.5
Gemini 4 Argon56.2
AA Analyst Agent AA Analyst Agent: Data-analyst tasks: the model explores data with tools and reports findings
Sonnet 5.5 · max57.5
Gemini 4 Argon48.8
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
Sonnet 5.543.9
Gemini 4 Argon57.1
Terminal-Bench Science 0.1 Terminal-Bench Science 0.1: Agents complete expert-authored research workflows across five scientific domains in a terminal
Sonnet 5.534.3
Gemini 4 Argon52.9
AA-Briefcase 1.1 · Elo AA-Briefcase 1.1 · Elo: Agents produce professional deliverables; blind pairwise comparisons produce the source Elo rating
Sonnet 5.51,635.8
Gemini 4 Argon1,485.8
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)
Sonnet 5.578
Gemini 4 Argon79.7
FrontierCode 1.1 FrontierCode 1.1: Agents implement changes in real repositories, graded against maintainer-defined requirements
Sonnet 5.544.3
Gemini 4 Argon43.5
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
Sonnet 5.5 · max2.8
Gemini 4 Argon3.9
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
Sonnet 5.5 · xhigh95.8
Gemini 4 Argon99.2
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
Sonnet 5.5 · xhigh70.6
Gemini 4 Argon92.2
Furniture Assembly Furniture Assembly: Models inspect furniture assembly photos, identify incorrect builds and locate the mistaken step
Sonnet 5.5 · max75
Gemini 4 Argon72.5
AutomationBench · Zapier run AutomationBench · Zapier run: End-to-end completion of business workflows using real SaaS applications
Sonnet 5.5—
Gemini 4 Argon51.3
DeepSWE 1.1 DeepSWE 1.1: Agents change repository code to resolve software engineering tasks under the stated DeepSWE edition
Sonnet 5.5 · max70.5
Gemini 4 Argon—
EBR Bench EBR Bench: Agents learn to play Earthborne Rangers across repeated playthroughs
Sonnet 5.5—
Gemini 4 Argon64.8
SpireClimb SpireClimb: Agents play Slay the Spire under the named game and difficulty setting
Sonnet 5.5 · max1.3
Gemini 4 Argon—
SpireClimb StS1 A20 Heart SpireClimb StS1 A20 Heart: Agents play Slay the Spire under the named game and difficulty setting
Sonnet 5.5 · max0
Gemini 4 Argon—
SpireClimb StS2 A10 SpireClimb StS2 A10: Agents play Slay the Spire under the named game and difficulty setting
Sonnet 5.5 · max2.7
Gemini 4 Argon—
| Benchmark | Sonnet 5.5 | Gemini 4 Argon |
|---|---|---|
| AutomationBench-AA | 59.4 | 77.5 |
| GDPval-AA (work tasks) 2.1 | 52.5 | 56.2 |
| AA Analyst Agent | 57.5 · max | 48.8 |
| Terminal-Bench 4.0 | 43.9 | 57.1 |
| Terminal-Bench Science 0.1 | 34.3 | 52.9 |
| AA-Briefcase 1.1 · Elo | 1,635.8 | 1,485.8 |
| AA-LCR (long context) 1.1 | 78 | 79.7 |
| FrontierCode 1.1 | 44.3 | 43.5 |
| Harvey LAB-AA 1.1 | 2.8 · max | 3.9 |
| SWE-bench Pro (v2 full) | 95.8 · xhigh | 99.2 |
| SWE-bench Pro (v2 hard) | 70.6 · xhigh | 92.2 |
| Furniture Assembly | 75 · max | 72.5 |
| AutomationBench · Zapier run | — | 51.3 |
| DeepSWE 1.1 | 70.5 · max | — |
| EBR Bench | — | 64.8 |
| SpireClimb | 1.3 · max | — |
| SpireClimb StS1 A20 Heart | 0 · max | — |
| SpireClimb StS2 A10 | 2.7 · max | — |
Specs
| Spec | Sonnet 5.5 | Gemini 4 Argon |
|---|---|---|
| Input, per 1M | $2 | !$2 |
| Output, per 1M | $10 | !$10 |
| Cost per task? | $0.88 | $1.99 |
| Speed, tok/s | 106 | 84 |
| Context | 1M | 1M |
| Weights | Closed | Closed |
| Released | Sep 2026 | Oct 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
Sonnet 5.5 → Gemini 4 Argon
Synchronous output stops at 65,536›
With background false, output limits above 65,536 tokens return HTTP 400; stay at or below that limit.
Sampling parameters are ignored›
temperature, top_p and top_k are ignored; remove them from requests.
No explicit caching›
Explicit caching is unavailable; omit explicit cache requests.
Gemini 4 Argon → Sonnet 5.5
No forced tool use›
Forced tool_choice any or tool returns HTTP 400; remove the forced choice.
Replace disabled thinking›
Disabled thinking returns HTTP 400; use between_tools for the lowest thinking setting.
No manual thinking budgets›
Manual thinking budgets and extra between_tools fields return HTTP 400; remove manual budgets and send between_tools without display, budget_tokens or block_binding.
Remove sampling parameters›
Non-default temperature, top_p or top_k returns HTTP 400; remove these parameters.
No assistant prefill›
A prefilled final assistant turn returns HTTP 400; remove that prefill.
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.