Models#Benchmark#Model inference
Google announces Gemini 4 Argon, skipping the promised 3.5 Pro — and you still cannot use it
Google announced Gemini 4 Argon on September 30: 77.9% on DeepSWE above GPT-6 Astra, a 1M-token output limit, phased release, no API pricing yet.

Google announced its Gemini 4 Argon flagship on September 30, jumping straight past the previously promised Gemini 3.5 Pro. The official numbers: 77.9% on the DeepSWE v1.1 coding benchmark, above GPT-6 Astra, Fable 5.1 and Opus 5.5, a 1M-token output limit, and chain-of-thought monitoring with stop systems. It is in a phased release — the general public cannot use it yet.
Facts
- Benchmarks: 77.9% on DeepSWE v1.1 per Google, above GPT-6 Astra, Fable 5.1 and Opus 5.5 (all official comparisons).
- Output limit: 1M tokens per response, up from 64K — aimed at long-generation tasks.
- Safety: chain-of-thought monitoring and stop systems, alongside a DeepMind “reasoning transparency” essay.
- Release cadence: “models of this scale call for a phased release” — trusted testers first (including Fairwind cybersecurity partners such as Wiz), then paid API plus AI Ultra, then enterprise, then consumers; no API pricing, no firm GA date.
- Internal claims: fleet-telemetry optimizations across 300 TiB of memory; 800K+ lines of C/C++ moved to Rust in the Fuchsia Zircon kernel.
- Third-party tests and pricing (October 1 update): first external results (Artificial Analysis, Gray Swan, Vals, Zapier, Harvey) give Argon 12 wins out of 18 benchmarks — DeepSWE v1.1 77.9% (Astra 74.1), Harvey legal agent 19.6% (Astra 5.4), prompt-injection attack success 0.7% (Astra 8.5%) — but it trails on long tasks and hallucination: FrontierSWE v2 55.0% (Astra 65.5%), Terminal-Bench 57.4% (Opus 66.4%), AA-Omniscience 15% vs Opus 59%; Artificial Analysis’ index of 53 ties Astra and sits below Opus 5.5. Pricing: $2/$10 per million tokens (in/out) at launch, then $4/$20.
Reading the third-party numbers
The first external data is best read as three contrasts: benchmark wins (12 of 18) coexist with a tied composite index (Artificial Analysis’ 53, level with Astra); a shrunken attack surface (0.7% prompt-injection success vs Astra’s 8.5%) coexists with worse hallucination (15% vs Opus’ 59%); a dominant legal vertical (Harvey 19.6% vs Astra’s 5.4%) coexists with weaker terminal tasks (Terminal-Bench 57.4% vs Opus’ 66.4%). The takeaway is not “which model wins” but that the strengths redistribute: writing and knowledge tasks converge on Argon, long-horizon terminal work stays with Opus.
Editorial take
Skipping version numbers and shipping-while-unavailable is the new flagship rhythm — OpenAI halted GPT-6.1 outright, Google pre-announces scores and trickles out access; different routes, same lesson: version numbers stopped being a capability calendar. Hold procurement math until launch pricing of $2/$10 reverts to $4/$20, and match the strength map above to your workload: test Argon for writing and knowledge-retrieval lines, keep GPT-6.1 Sol’s pricing and Opus for terminal automation.
Access path and timeline
- Who gets it first: Fairwind cybersecurity partners such as Wiz, then paid API plus AI Ultra, then enterprise, then consumers; The Verge adds that Google has joined the federal government’s voluntary pre-release evaluation process.
- Pricing: $2/$10 per million tokens (in/out) at launch, reverting to $4/$20 — against GPT-6.1 Sol’s one-fifth-of-Astra pricing, the flagship price war is officially on.
- Internal use: Google says Argon already powers its own large-scale codebase migrations, consistent with the 800K-line Rust migration claim.
Relation to the previous generation
The Gemini 3.5 Pro version number was skipped with no explanation; shipped Gemini 3 models remain in service. Teams that buy by version number should switch to a triple of benchmark mix, pricing and availability dates when evaluating model generations — the shared lesson of this cycle’s flagship moves (Argon skipping a number, GPT-6.1 scrapped, Sol slotting in early).
One more note for evaluation pipelines: the external results came from named labs (Artificial Analysis, Gray Swan, Vals) plus application vendors (Zapier, Harvey), not from Google’s own harness. That is a healthier mix than launch-day self-benchmarks, and it arrived while the model is still gated — a sequencing worth demanding from every lab.