Claude Sonnet 5 vs Qwen 3.7 Max: The Coder vs The Marathon Runner

Two models launched 41 days apart. Claude Sonnet 5 — Anthropic's new mid-tier king, June 30, 63.2% Pro, $3/$15. Qwen 3.7 Max — Alibaba's "Agent Frontier," May 19, 60.6% Pro, $1.25/$3.75. On shared benchmarks, Sonnet leads by slim margins (+2.6 Pro, +4.8 Verified, +1.8 HLE). But Qwen fights back with world-class math (GPQA 92.4%, HMMT 97.1%), proven 35-hour autonomous operation, and a price tag that's 2.7× cheaper on output. One model writes your code. The other runs marathons. Here's the complete comparison, sourced from Anthropic's Sonnet 5 System Card and Qwen's official blog. Test both on CodingFleet.

Head-to-Head: Shared Benchmarks

Sonnet 5 from Anthropic System Card. Qwen 3.7 Max from Qwen Official Blog and DataCamp analysis. Both vendor-reported. Cross-vendor comparisons are directional — different scaffolds introduce uncertainty. Terminal-Bench excluded: Sonnet uses TB 2.1, Qwen uses TB 2.0 (different task sets, not comparable).

Coding: Sonnet's Modest Edge

On SWE-bench Pro — the benchmark that matters most for production coding — Sonnet 5 leads 63.2% to 60.6%. That's a 2.6-point gap. On SWE-bench Verified (the classic 500-problem set): Sonnet at 85.2% vs Qwen at 80.4% (+4.8). On HLE without tools (raw reasoning): Sonnet at 43.2% vs Qwen at 41.4% (+1.8).

These are real leads, but they're not decisive. A 2.6-point Pro gap on 1,865 tasks is ~48 tasks. For most development workflows, both models will handle the vast majority of coding tasks competently. Sonnet gets the edge on the hardest bugs; Qwen gets the edge on affordability.

The Terminal-Bench comparison is unfortunately apples-to-oranges. Sonnet 5 scores 80.4% on Terminal-Bench 2.1 (the current standard, same harness used for Opus 4.8 and GPT-5.5). Qwen 3.7 Max scores 69.7% on Terminal-Bench 2.0 (an older, different task set). The version gap makes direct comparison unreliable. DataCamp's analysis: "This benchmark tests autonomous terminal-based software engineering with a 5-hour timeout and 12 CPU cores. On SWE-Pro, it scores 60.6, the highest in the comparison table."

Math: Qwen's Dominance

This is where the comparison flips. Qwen 3.7 Max is a math powerhouse:

GPQA Diamond at 92.4% is elite — competitive with Opus 4.8 and GPT-5.5. HMMT at 97.1% and AIME at 98.3% are ceiling scores. For math-heavy coding (scientific computing, algorithmic work, formal verification), Qwen 3.7 Max has a clear advantage. Sonnet 5's USAMO score (79.5%) is impressive in its own right, but Anthropic didn't publish GPQA Diamond or competition math benchmarks for this model — they focused the System Card on agentic coding and safety instead.

Agentic Autonomy: Qwen's 35-Hour Marathon

This is Qwen's signature capability — and it's something no other model in this class can claim. Yotta Labs' coverage of Qwen's launch:

"A 35-hour autonomous kernel optimization run on T-Head ZW-M890 PPUs. Across 432 kernel evaluations and 1,158 tool calls, the model finished with a 10x geometric mean speedup. Most agent models stop making progress after a few hours. Qwen 3.7-Max sustained meaningful progress past 30 hours."

And during an 86-hour RL training session, the model autonomously flagged 1,618 reward hacking cases and added 13 new heuristic rules to its own training loop. Kernel Bench L3: 96% win rate. This is the "Agent Frontier" label in action — Qwen 3.7 Max is built for unattended, long-horizon autonomous operation.

Sonnet 5's agentic capabilities are strong (80.4% TB 2.1, 81.2% OSWorld), but Anthropic hasn't demonstrated the kind of multi-hour autonomous endurance that Qwen has proven. If your workflow involves overnight agent runs or continuous autonomous optimization, Qwen's track record matters.

Pricing: Qwen is 2.7× Cheaper on Output

For a workload of 10M input + 1M output tokens per day:

Assumes 90% input / 10% output split, 0.5× thinking multiplier, no caching. Qwen promotional pricing.

At $488/month for Qwen vs $1,350/month for Sonnet 5 standard, the annual difference is ~$10,300. That's real money for teams running production agents.

The Radar: Coding vs Math + Autonomy

Where They Don't Overlap

Specification Comparison

*Qwen 3.7 Max promotional pricing: 50% off through June 22, 2026. Standard pricing: $2.50/$7.50. Both OpenAI and Anthropic API compatible. Sources: Qwen Official Blog, Yotta Labs.

Should You Use Sonnet 5 or Qwen 3.7 Max?

Conclusion: The Coder vs The Marathon Runner

Claude Sonnet 5 and Qwen 3.7 Max represent two different answers to the question "what should a near-frontier model excel at?" Sonnet 5 is the better coder — +2.6 Pro, +4.8 Verified, deep Anthropic ecosystem integration, and the safety infrastructure that enterprises trust. It's the model you want reviewing your pull requests.

Qwen 3.7 Max is the better autonomous agent — 35-hour proven endurance, world-class math, 2.7× cheaper, and OpenAI/Anthropic API compatible. It's the model you want running overnight optimization jobs and tackling algorithmic heavy lifting.

Qwen's positioning: "The Agent Frontier." Anthropic's positioning: "The best combination of speed and intelligence." Both true. The 2.6-point Pro gap makes Sonnet the better coder. The 2.7× price gap makes Qwen the better value. Your choice depends on whether you value marginal coding quality or substantial cost savings — and whether you need a model that writes code or one that runs marathons.

Sources & Links

• Anthropic — Claude Sonnet 5 System Card — Table 8.1.A capability evaluation summary
• Anthropic — Introducing Claude Sonnet 5 — official launch announcement
• Qwen Official Blog — Qwen 3.7 Max — benchmark comparison table, Agent Frontier claims
• DataCamp — Qwen 3.7 Max: Features, Benchmarks, Agent Capabilities
• Yotta Labs — Qwen 3.7 Max: Pricing, Features, Access
• CodingFleet — GLM-5.2 vs Qwen 3.7 Max
• Pioneer AI — How to Choose the Best Coding Models (2026)
• Claude Platform Docs — Models Overview