Terminal Execution (TB 2.1)
90.8%
Gemini 3.8 Flash (+22.3% vs 3.1 Pro)
Abstract Logic (ARC-AGI-2)
77.1%
Gemini 3.1 Pro (Frontier Ceiling)
Inference Throughput
313 tps
Gemini 3.8 Flash (0.70s TTFT at Low)
SWE-bench Verified
80.8%
Claude Opus 4.6 (3.1 Pro: 80.6%)
Benchmark Profiles Across Evaluation Suites
Evaluating SWE-bench Verified, Terminal-Bench, ARC-AGI-2, and DeepSWE v1.1.
Throughput (TPS) vs Input Cost ($/1M Tokens)
Speed-to-cost efficiency curve for Antigravity iterative development.
Comprehensive Four-Model Antigravity Evaluation
| Model | Role in Antigravity | Input / Output ($/1M) | TTFT & Speed | SWE-bench / Terminal | Abstract Reasoning |
|---|---|---|---|---|---|
| Gemini 3.8 Flash | Execution Workhorse | $0.75 / $3.75 (promo) | 0.70s – 13.3s | 313 tps | 89.4% (scaff) | 90.8% (TB 2.1) | 45.4% – 52.0% (Tactical) |
| Gemini 3.1 Pro | Logic & Spec Anchor | $2.00 / $12.00 (≤200k) | 27.9s – 44.5s | 112 tps | 80.6% | 68.5% (TB 2.0) | 77.1% (ARC-AGI-2) |
| Claude 4.6 Sonnet | Balanced Editor | $3.00 / $15.00 | ~0.96s | 45–55 tps | 79.6% | 59.1% (TB 2.0) | 56.4% (OSWorld 72.5%) |
| Claude 4.6 Opus | Senior Architect | $5.00 / $25.00 | 2.5s – 45.0s | 25–35 tps | 80.8% | 62.4% (TB 2.0) | 68.8% (Design Master) |
When to Use Gemini 3.1 Pro
- Out-of-Distribution Math & Abstract Proofs: When writing custom algorithms, compilers, or serialization specs where ARC-AGI-2 reasoning (77.1%) prevents false lemmas.
- Dense Long-Context Synthesis (>512k Tokens): In large multi-service codebases, 3.1 Pro preserves semantic grounding across 1M tokens without "lost-in-middle" attention dispersion.
- Ambiguity Disambiguation: Halts execution to surface underspecified requirements rather than prematurely committing code based on shaky assumptions.
- Massive Output Blocks: Emits sustained multi-file codebases up to 65,536 output tokens with global coherence.
When to Use Claude 4.6 Opus / Sonnet
- Greenfield Architectural Decomposition: Claude Opus 4.6 acts as Senior Architect, designing clean abstractions and module boundaries without speculative boilerplate.
- Strict Repository Conventions: When an explicit `CLAUDE.md` or architecture specification exists, Opus follows constraints with military precision.
- Subtle Multi-File Refactoring: Guarantees zero regressions in complex type systems (Rust, TypeScript, Go interfaces) where localized greedy edits would fail.
- GUI & Desktop Orchestration: Claude Sonnet 4.6 leads in OSWorld-Verified (72.5%) for interactive browser/GUI manipulation.
Adversarial Red-Team Analysis: Critical Failure Modes of Gemini 3.8 Flash
The "Work Harder" Cost Paradox: 3.8 Flash achieves high success rates by running 10–15 iterative bash/eval tool cycles. Because thinking tokens are billed at full output rates ($3.75/1M), a protracted retry storm on an ambiguous bug can consume more tokens and wall-clock time than a single-pass 3.1 Pro or Claude Opus invocation.
Myopic Greedy Patching: Flash models optimize for the immediate test or compiler error. In complex refactors, this frequently causes "whack-a-mole" regressions where fixing one unit test silently breaks cross-package invariants.
Deprecation of 'Minimal' Thinking: Gemini 3.8 Flash removes the minimal thinking mode entirely (triggering API validation errors if passed). Every invocation carries mandatory reasoning token overhead.
Recommended Antigravity Dispatch Architecture
| Workflow Phase | Assigned Model | Thinking Level | Key Rationale |
|---|---|---|---|
| 1. Architecture & Blueprint Planning | Claude Opus 4.6 or Gemini 3.1 Pro | `high` / `adaptive` | Prevents greedy code sprawl, preserves global dependency contracts, and resolves ambiguous specs. |
| 2. Iterative TDD & Terminal Debugging | Gemini 3.8 Flash | `medium` | 90.8% Terminal-Bench 2.1, sub-second TTFT, fast test execution cycles, 313 tps generation. |
| 3. Asynchronous Subagent Swarms | Gemini 3.8 Flash | `low` | Spawns 4–8 concurrent research and lint workers without blowing token-per-minute rate limits. |
| 4. Whole-Repo Synthesis (>500k tokens) | Gemini 3.1 Pro | `medium` | Dense MoE attention prevents lost-in-the-middle degradation during repo-wide audits. |
| 5. Automated CVE & Vulnerability Patching | Gemini 3.8 Flash Cyber | `Fairwind Mode` | Specialized security harness; 47.2% CWE-Bench autonomous patch pass rate. |