Google started widening Gemini 4 Argon availability on October 13, after the September 30 debut that reached cyber-defense partners first. Developers and enterprises now get access through AI Studio, the Gemini API, Antigravity, and Gemini Enterprise.

Launched: Sep 30, 2026 → expanded rollout Oct 13
Intro: $2 input / $10 output per million tokens (until Nov 30)
List: $4 / $20 after Nov 30
Cached input: 95% off, sustained

Price matters, but the scoreboard matters first, because it splits by exam.

Two report cards: first on Vals, behind on terminal tasks

From the DeepMind model card and Google's announcement:

Vals Index (finance, law, tax, coding composite)
  Argon 68.9% > Opus 5.5 67.0% > Astra 63.1% (1st)

DeepSWE v1.1 (real multi-step engineering)
  Argon 77.9% > Opus 5.5 74.2% > Astra 74.1% (1st)

Vals Finance Agent v2
  Argon 65.4% > Fable 5.1 58.9% > Opus 5.5 58.6% (1st)

Harvey Legal Agent
  Argon 19.6% vs Astra 5.4% (wide lead)

AutomationBench 51.3%, LVBench 91.7%, CWE-bench 68% (tied 1st)

And where it flips:

Terminal-Bench 4.0 (multi-step terminal execution)
  Opus 5.5 66.4% > Astra 58.2% > Fable 5.1 57.9% > Argon 57.4%

FrontierSWE v2
  Astra 65.5% > Opus 5.5 62.3% > Argon 55.0%

Terminal-Bench Science 0.1
  Astra 68.1% > Opus 5.5 63.3% > Argon 57.6%

Long-horizon engineering plus finance and legal knowledge work: ahead. Sharp terminal execution: rivals lead. First overall, not first everywhere.

This is the same structure our SWE-bench vs Terminal-Bench vs ProgramBench guide describes. Different exams, so don't add the scores.

1M output tokens means long jobs, not long essays

Argon's signature is unchanged: per-response output ceiling up from 64K to 1M tokens.

Context (reading)
  "How much material fits at once?"

Output (thinking, tools, writing)
  "How long can one work trajectory run?"

Google's internal cases show the direction: an 800K-line Fuchsia Zircon C/C++ to Rust migration, a libgav1 Rust SIMD rewrite at 2.7x speedup, all with automated checks, emulation tests, and human review alongside. Not a model working alone.

One caution travels with it: analyses put average Vals Index task cost near $15.68, code migration near $58, and CUA-bench near $194. Models that work longer bill longer.

Budget at list price, not launch price

The $2/$10 intro ends November 30, then $4/$20, the Opus 5.5 tier.

Now: $2 / $10 (same as Sonnet 5.5, GPT-6.1 Sol)
After Nov 30: $4 / $20

Run comparisons at $4/$20 from day one, or December's bill surprises you.

CodeBridge mini lab: run three models on one job

1. Fix one job (e.g. fixing five failing tests)
2. Keep harness, tools, and time limits identical
3. Run Argon / GPT / Claude and record:

   [ ] completion rate (test-pass criterion)
   [ ] p95 latency (not the mean)
   [ ] cost per success (at $4/$20 list price)
   [ ] retries and human interventions
4. Split results by exam type (DeepSWE-style vs terminal-style)

Bring the lesson from our million-token context cost guide: stuffing long inputs every time inflates cost even with cache discounts. Feeding only what is needed matters more on long-horizon models like Argon.

Conclusion: pick the exam, not just the model

Long refactors, finance, legal knowledge work → compare Argon first
Terminal execution and environment wrangling → include Opus and Astra
Pricing → compute at December rates

Completion rate, p95, and cost per success on your work decide. Run one repo job through all three models identically this week; that table picks next quarter's tooling.

Further reading

References

Go deeper with a course

Long jobs are finished by structure, not by the model alone. Practice placing memory and verification inside harness, loop, and graph designs.