Google's Gemini 3.6 Flash Scores 49% on Coding Benchmark While Cutting Costs 52%
Gemini 3.6 Flash hits 49% on DeepSWE — matching Claude Opus 4.8 at medium reasoning — while using 65% fewer tokens and costing 52% less per task.
Takeaways
- 49% on DeepSWE: Gemini 3.6 Flash scores 49% (up from 37%), matching Claude Opus 4.8 at medium reasoning effort.
- 65% fewer tokens: On long-horizon coding tasks, 3.6 Flash uses up to 65% fewer output tokens than its predecessor, cutting cost per task by 52%.
- New pricing: $1.50/1M input tokens and $7.50/1M output tokens — down from $9.00/1M output on 3.5 Flash.
- Broader gains: MLE-Bench jumps from 49.7% to 63.9%; OSWorld-Verified (computer use) hits 83.0%, best in class.
- 1M context, updated knowledge: Both 3.6 Flash and 3.5 Flash-Lite share a 1M-token context window and a March 2026 knowledge cutoff.
- Available now: Accessible via the Gemini API and Google AI Studio; full leaderboard at DeepSWE.
Google just released Gemini 3.6 Flash, and the headline number is deceptively simple: 49% on DeepSWE, up from 37% for its predecessor. The real story is that the model achieved that score while spending dramatically fewer tokens to get there. That combination is rare, and it changes the math for anyone running coding agents at scale.
What landed on the leaderboard
Google launched three new Gemini models aimed at making AI agents faster, cheaper, and more efficient: Gemini 3.6 Flash, Gemini 3.5 Flash Lite, and Gemini 3.5 Flash Cyber. Gemini 3.6 Flash is the flagship of the three and the one that matters most for coding agent workloads.
The model scored 49% on DeepSWE, compared with 37% for Gemini 3.5 Flash, a 12-point absolute improvement on a benchmark specifically designed to be hard to game. Higher precision and fewer unnecessary code changes drove the gain.
What DeepSWE actually measures
Most public coding benchmarks are saturating. Top models cluster within a narrow score band where differences fall inside the margin of error. DeepSWE was built to fix that:
- Contamination-free: Tasks are written from scratch, not adapted from existing commits or PRs, so no model has seen the solution during pretraining.
- High diversity: 113 tasks across 91 repositories in 5 languages: TypeScript, Go, Python, JavaScript, and Rust.
- Real-world complexity: Prompts run roughly half the length of SWE-bench Pro's, yet reference solutions average 668 lines of code added across 7 files, 5.5x more code than SWE-bench Pro tasks.
- Reliable verification: Verifiers are hand-written to test observable behavior, not implementation details. SWE-bench Pro's verifier misgrades outputs at 8% false positives and 24% false negatives; DeepSWE's rates are 0.3% and 1.1%.
The benchmark runs every model through mini-swe-agent, a model-agnostic harness that gives every model the same bash tool and system prompt. No per-vendor editing primitives, no model-specific tuning. This isolates model capability from scaffolding effects.
Efficiency is the real story
A score jump from 37% to 49% is meaningful. What makes Gemini 3.6 Flash genuinely interesting for production use is how it got there:
- On the Artificial Analysis Index, 3.6 Flash uses 17% fewer output tokens than 3.5 Flash. On DeepSWE, Google reports up to a 65% reduction.
- The model takes fewer reasoning steps and tool calls per multi-step workflow.
- Pricing drops to $1.50 per 1M input tokens and $7.50 per 1M output tokens, down from $9.00 per 1M output tokens on 3.5 Flash.
On the DeepSWE leaderboard, Gemini 3.5 Flash averaged $7.34 per task with 276k output tokens. Datacurve's numbers put the efficiency gains from 3.6 Flash at a 52% cost reduction per agentic task. Higher score, fewer tokens, lower cost per run.
Where it sits in the competitive landscape
The 49% score puts Gemini 3.6 Flash roughly on par with Claude Opus 4.8 at medium reasoning effort, a much heavier and more expensive model. That's the competitive framing Google is leaning into: Flash-tier speed and price with performance that previously required flagship-tier models.
The full DeepSWE v1.1 leaderboard shows the top of the board occupied by models scoring in the 67–73% range:
- gpt-5.6-sol [max]: 73% at $8.39/task
- claude-fable-5 [max]: 70% at $21.63/task
- gpt-5.6-terra [max]: 70% at $4.95/task
- kimi-k3 [max]: 69% at $4.65/task
- gpt-5.5 [xhigh]: 67% at $7.23/task
GPT-5.6 Luna leads on DeepSWE (67%) and Terminal-bench (84.7%). Grok 4.5 edges ahead on SWE-Bench Pro (64.7%). Claude Sonnet 5 tops both MLE-Bench (66.9%) and GDPVal-AA v2 (1607). Gemini 3.6 Flash holds the top spot on OSWorld-Verified, both CharXiv Reasoning variants, and both GDM-MRCR long-context tests. No single model sweeps every category.
Beyond coding: the broader benchmark picture
The improvements extend beyond software engineering tasks:
- MLE Bench score climbed to 63.9% from 49.7%.
- OSWorld-Verified score reached 83.0%, up from 78.4%, with computer use available as a built-in client-side tool via the Gemini API and Gemini Enterprise.
- On GDM-MRCR v2 at 1M tokens (pointwise), Gemini 3.6 Flash scores 54.0% while Gemini 3.5 Flash and Gemini 3.1 Pro both sit below 27%.
- The knowledge cutoff advances from January 2025 to March 2026.
What this means for agentic pipelines
If you're running a coding agent in production, whether for automated PR review, issue resolution, or multi-step refactoring, Gemini 3.6 Flash is now a serious option where it previously wasn't. Token efficiency matters especially in long-horizon tasks where agents loop through many tool calls. Fewer steps per task means lower latency and lower cost, compounding across thousands of runs.
The model also delivers higher precision with fewer unwanted code edits and reduced execution loops compared to its predecessor. Agents that over-edit produce noisy diffs that are harder to review and more likely to introduce regressions, so that property has real downstream value.
Gemini 3.6 Flash is available now through the Gemini API and Google AI Studio. Both Gemini 3.6 Flash and Gemini 3.5 Flash Lite carry a 1-million-token input context window, a 64,000-token output limit, and a knowledge cutoff of March 2026.
The Flash tier is no longer a budget compromise. The DeepSWE numbers make a credible case that efficiency and capability can move in the same direction at once.