Gemini 3.7 Flash (Google DeepMind, August 13, 2026) and Claude Sonnet 5 (Anthropic, June 30, 2026) are the two mid-tier workhorses most teams will actually run at volume this year. One was built for raw speed; the other for agentic depth, computer use, and safety. This is the direct sequel to our Gemini 3.6 Flash vs Claude Sonnet 5 comparison from July — and a lot has changed in six weeks.

This guide compares the two across benchmarks, pricing, speed, context, and modalities. Figures are drawn from Google's launch chart, Anthropic's documentation, Artificial Analysis, and independent coverage. It's part of our Gemini 3.7 Flash series — see also Grok 4.6 vs Gemini 3.7 Flash.

TL;DR: Gemini 3.7 Flash now beats Sonnet 5 on the AA Intelligence Index (56 vs 55) while generating ~5× faster (340 vs 69 t/s) and costing 5.7× less per completed task ($0.40 vs $2.29). But Sonnet 5 retains three real advantages: computer use (Agent's Last Exam 33.3% vs 26.3%), deeper knowledge work (GDPVal-AA 1598 vs 1525), and the Claude Code ecosystem. Google wins the efficiency war; Anthropic still wins on desktop agents and safety posture.

At a Glance

CategoryGemini 3.7 FlashClaude Sonnet 5
Release dateAugust 13, 2026June 30, 2026
DeveloperGoogle DeepMindAnthropic
Context window1,048,576 tokens1,000,000 tokens
Max output65,536 tokens128,000 tokens
Input modalitiesText, image, video, audio, PDFText, image (+ computer use)
Input price$0.75 / 1M (intro)$2 / 1M (permanent)
Output price$3.75 / 1M (intro)$10 / 1M (permanent)
Price after intro$1.50 / $7.50 (Jan 1, 2027)No change — $2/$10 locked in
ReasoningLow / medium / highAdaptive thinking, 5 effort levels (xhigh)
Model IDgemini-3.7-flashclaude-sonnet-5

Benchmark Comparison

Google's own launch chart is unusually candid: Sonnet 5 wins some rows. Here's where each model leads.

EvaluationGemini 3.7 FlashClaude Sonnet 5Leader
AA Intelligence Index5655 (max)Gemini 3.7
GDPVal-AA v2 (Elo)15251598Sonnet 5
AutomationBench30.4%10.7%Gemini 3.7
GDP.pdf34.0%28.0%Gemini 3.7
FrontierCode 1.1 Main43.6%42.7%≈ Tie
Agent's Last Exam (desktop/OS)26.3%33.3%Sonnet 5
OSWorld-Verified81.2%Sonnet 5
BrowseComp84.7%Sonnet 5
Terminal-Bench85.8% (2.1)80.4% (2.0)Version mismatch

Sources: Google DeepMind launch chart (Aug 2026), Anthropic system card, Artificial Analysis model pages. Terminal-Bench rows use different versions (2.1 vs 2.0) and are not directly comparable — see our Terminal-Bench 2.1 leaderboard for apples-to-apples data.

Agentic Benchmarks Head-to-Head

The AutomationBench gap is the story of this matchup: Gemini completes real enterprise workflows nearly 3× as often as Sonnet 5 (30.4% vs 10.7%) per Google's chart. Gemini also edges GDP.pdf document comprehension (34.0 vs 28.0) and effectively ties FrontierCode (43.6 vs 42.7). Sonnet 5's counter: Agent's Last Exam, the multimodal desktop-and-OS evaluation, where it leads by 7 points — and its OSWorld-Verified 81.2% remains among the best computer-use scores published.

Knowledge-Work Elo

Sonnet 5's GDPVal-AA Elo of 1598 leads Gemini's 1525 by 73 points — consistent with the pattern in our Claude Opus 5 analysis, where Anthropic models dominate knowledge-work Elo. If your workload is professional documents, structured analysis, and judgment-heavy deliverables, Sonnet 5 remains the stronger of these two — but note it pays for that depth: at max effort Sonnet 5 uses ~40% more output tokens per task than its predecessor.

Capability Radar

Radar values are normalized (0–100) to the higher of the two models on each axis: Intelligence = AA Index, Knowledge Work = GDPVal-AA v2, Automation = AutomationBench, Documents = GDP.pdf, OS/Computer = Agent's Last Exam, Speed = output t/s.

Pricing: Gemini's Discount vs Sonnet's New Permanent Rate

Update (August 14, 2026): Anthropic has cancelled Sonnet 5's planned September 1 price increase — its $2/$10 launch rate is now permanent. That changes the calculus: Gemini's $0.75/$3.75 is still a limited-time discount (through December 31, 2026, then $1.50/$7.50), while Sonnet's $2/$10 has no expiry date at all. Cost per completed task still favors Gemini:

Cost componentGemini 3.7 FlashClaude Sonnet 5
Input (per 1M)$0.75 (intro)$2 (permanent)
Output (per 1M)$3.75 (intro)$10 (permanent)
Cost per AA Index task$0.40$2.29 (max)

Consider a hypothetical agent task using 1M input + 250K output tokens:

That task costs roughly $1.69 on Gemini vs. $4.50 on Sonnet 5 — 2.7× cheaper. The per-task gap is larger still: $0.40 vs $2.29 (5.7×), because Sonnet 5's max effort burns far more tokens per task. One nuance worth knowing: Sonnet 5 uses an updated tokenizer that produces ~1.0–1.35× more tokens than Sonnet 4.6 for the same input — so its real per-character cost is higher than the sticker suggests.

Speed & Latency

No contest on decode speed — Gemini is the fastest reasoning model on the market:

MetricGemini 3.7 FlashClaude Sonnet 5
Output speed~340 t/s (high)69.4 t/s (max) / 80 t/s (xhigh)
Time to first token~2.7 s (AIHubMix, traffic)202 s (max) / ~2 s (low effort)
Time per AA Index task1.7 min

Gemini streams ~5× faster than Sonnet 5 at max effort (340 vs 69 t/s). Sonnet 5's max-effort TTFT of ~202 seconds is one of the longest of any frontier model — a direct consequence of adaptive thinking that runs very long before answering. At low effort Sonnet 5 is near-instant, but then you lose the intelligence. If your product has interactive UX, this latency profile is Sonnet's biggest practical weakness; for batch agent jobs that run for hours, it matters far less.

Context & Modalities

Both ship ~1M-token context windows with no long-context surcharge. The differences are at the edges:

  • Max output: Sonnet 5 doubles Gemini (128K vs 65,536 tokens) — better for long code generations and full-document rewrites.
  • Multimodality: Gemini accepts video and audio; Sonnet 5 is text + image but adds computer use — controlling the desktop, browsers, and GUIs, where its OSWorld-Verified and Agent's Last Exam scores lead this matchup.
  • Ecosystem: Sonnet 5 is integrated into Claude Code and is the default for Claude Free/Pro. Gemini 3.7 Flash lives in the Gemini API, AI Studio, Antigravity, and Gemini Enterprise.
  • Safety: Anthropic ships Sonnet 5 with real-time cyber safeguards enabled by default and reports lower hallucination/sycophancy rates than Sonnet 4.6.

Verdict: Which Should You Choose?

Choose Gemini 3.7 Flash if you're optimizing for speed, cost, or multimodal input: high-volume agents, video/audio pipelines, and interactive apps. It is now smarter than Sonnet 5 on the AA Index while costing 5.7× less per task and streaming 5× faster — but unlike Sonnet's new permanent $2/$10, Gemini's discount ends December 31, 2026.
Choose Claude Sonnet 5 if you need desktop agents, Claude Code tooling, long outputs, or judgment-heavy knowledge work: computer use (Agent's Last Exam +7), GDPVal-AA (+73 Elo), 128K output, and Anthropic's safety posture. Budget for its token hunger — at max effort it's one of the most token-hungry models in its class.

The honest takeaway from Google's own chart: Gemini 3.7 Flash doesn't sweep — it wins the volume lanes (automation, documents, speed, price) while Sonnet 5 holds the desktop-agent and knowledge-work lanes. Many teams will run both: Gemini as the default high-throughput model, Sonnet 5 for computer-use and Claude Code workloads. If you're coming from our Grok 4.6 vs Gemini 3.7 Flash analysis: Grok beats both on raw intelligence (61), but neither Sonnet nor Gemini can touch Gemini's speed-per-dollar.

Sources