Gemini 4 Argon's benchmark win is real. The coding narrative needs a footnote.
Jump to a section
See the story at a glance
If you use any of Google's AI tools — Gemini in Search, in Google Docs, in Workspace — something changed last Tuesday. Google announced Gemini 4 Argon on September 30, its first new frontier model in months. The announcement led with a headline number: 77.9% on DeepSWE v1.1, ahead of Claude Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%.
The number is real. Two other coding benchmarks don't look as good for Google. That split is the story.
Source spread
- Google DeepMind — Gemini 4 Argon — hype. Official company announcement. 19-benchmark comparison table, pricing, access tiers. All numbers first-party; no independent methodology links cited.
- TechCrunch — Google releases Gemini 4 Argon — builder. Straightforward launch coverage; notes the Fairwind-first access restriction and the government voluntary review process.
- benchlm.ai — Gemini 4 Argon benchmark analysis — skeptic. Aggregates the full benchmark picture, including Terminal-Bench and FrontierSWE v2, which Google did not lead with.
What's real
The knowledge-work benchmarks are genuinely strong. On legal automation (Harvey's Legal Agent Benchmark: Argon 19.6% versus next model's 6.7%) and finance (Vals Finance Agent v2: Argon 65.4%), Argon leads by a margin that isn't close. These benchmarks test what happens when an AI has to read large amounts of documents and make decisions across them — the kind of thing lawyers and analysts do all day. The 1M output token limit matters here. Think of the output limit like a whiteboard: a much bigger whiteboard means the AI can return more work in one go before you have to start clearing it. One million tokens is a real capability jump for long documents and multi-step tasks.
CWE-bench v1 at 68%, tied for first with GPT-6 Astra: the one security benchmark where Google isn't alone in first. The Fairwind-first rollout — to vetted cyber defenders before anyone else — matches this positioning deliberately.
Pricing is competitive at $2/$10 per million input/output tokens intro rate. The 95% discount on cached input makes repeated queries against a fixed large document very cheap.
What deserves a side-eye
FrontierSWE v2: Argon 55.0%, Claude Fable 5.1 at 56.3%, Claude Opus 5.5 at 62.3%, GPT-6 Astra at 65.5%. Argon finishes last on this one. FrontierSWE tests real software engineering — fixing bugs and implementing features in existing codebases, not structured problem sets. Ten points behind the leader.
Terminal-Bench 4.0: Argon 57.4%, Claude Opus 5.5 at 66.4%. Terminal-Bench drops an agent into a real shell and asks it to complete tasks — closer to how most developers actually use these models. Google's own comparison table includes this row. Argon loses it by nine points.
The gap between DeepSWE (Argon wins by 3 points) and Terminal-Bench (Argon loses by 9 points) is the kind of thing that should make you pause before accepting the "best at coding" framing. Both are coding benchmarks. They test different things: structured problem-solving versus sustained autonomous operation in a real environment.
Also: Google's 19-benchmark table has no third-party verification at launch. These numbers come from Google. Standard for a model announcement, and worth keeping in mind.
| Benchmark | Gemini 4 Argon | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| DeepSWE v1.1 | 77.9% | 74.2% | 74.1% |
| FrontierSWE v2 | 55.0% | 62.3% | 65.5% |
| Terminal-Bench 4.0 | 57.4% | 66.4% | — |
| CWE-bench v1 | 68% | 67% | 68% |
What to do about it
For everyday Gemini users: Nothing changes today. Argon launched to vetted cybersecurity programs only — not yet to Google Search, Docs, or Workspace. When it reaches consumer products, you'll most likely notice it on long, complex tasks: summarizing a big document, drafting a detailed report, multi-part research questions. Keep an eye on Google's product announcements for the broader rollout.
For builders evaluating models:
- Don't switch based on DeepSWE alone. That benchmark and Terminal-Bench tell different stories about the same model.
- If your workload is long-document processing, legal or finance automation, or large-context knowledge retrieval: Argon is worth testing at the intro price.
- If your workload is terminal-heavy agent tasks: run your own eval before switching. The nine-point Terminal-Bench gap matters.
- Wait for third-party benchmark reproduction before making production decisions.
- Factor in the post-intro pricing ($4/$20/M) when doing production cost math.
Further reading
- Google DeepMind — Gemini 4 Argon — all first-party benchmarks, pricing, and access details
- TechCrunch — Google releases Gemini 4 Argon — launch news with access and government review context
- benchlm.ai — Gemini 4 Argon — aggregated benchmark comparisons including Terminal-Bench and FrontierSWE v2
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Google opened Carbon to everyone. The price doubles in three weeks.
Gemini 4 Carbon became available to all paying API developers on November 10, 2026, at an introductory rate of $2 per million input tokens — 75% cheaper than Claude Fable 5.2. The intro rate expires November 30. Terminal-Bench 4.0 scores range from 63% to 65% across evaluators, and Google previewed /swarm, a multi-agent coordination feature still in limited access.
Fable 5.2 leads the coding benchmarks. The science numbers depend on who's running them.
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on November 9 at $8/$40 per million tokens — 20% cheaper than Fable 5.1 and GPT-6 Astra. On coding and agentic benchmarks the lead holds across evaluators. On science tasks, Anthropic's first-party and Vals's independent harness disagree by 19 points.
Claude Haiku 5.5 stopped being the compromise option
Claude Haiku 5.5 landed October 7 at $0.10/M input tokens — a tenfold cut from Haiku 4.5's $1.00/M. Terminal-Bench 4.0 went from 0% to 39.2%. The model tier you routed around is worth reconsidering.