Google launches Gemini 4 Argon, its first frontier model in seven months, gated to cyber defenders first

For a CTO in Dubai or Riyadh, the number worth quoting is $0.74. Artificial Analysis measured that as cost per completed task for Gemini 4 Argon (high). GPT-6 Astra (max) runs $1.21 at the same Intelligence Index of 53. Claude Opus 5.5 is $2.70. Google's pricing sits at $2 per 1M input and $10 per 1M output during the introductory period, then $4/$20. Cached input is $0.10 per 1M. Add the 1M-token output ceiling, up from 64K and roughly eight times the 128K that Astra and Opus 5.5 offer. Whole-repo migrations become cheaper per finished job than anything currently public. The catch is access. Fairwind members first. Enterprise and Vertex AI "in the coming weeks." No date. Plan roadmap and procurement around access windows, not launch days.

Treat vendor benchmarks as unaudited. Google reports 76.2% on SWE-Bench Pro, 61.6% on Terminal-Bench 2 and 68% on CWE-Bench for vulnerability remediation. Those run on a proprietary scaffold. Independent figures come in at 62% and 64% on the first two. The Decoder shows Argon trailing Claude Opus 5.5 on Humanity's Last Exam (49.2% against Astra's 57.3%, and Astra itself behind Opus on GAP-40) and on ARC-AGI-2 (22.1% against 46.1%). The 14-point gap between Google's SWE-Bench Pro score and independent measurement is larger than most gaps between competing models. Its 15% hallucination rate on Hal Bench v2 is a real improvement on Astra's 51%. It is not, however, a licence to let agents publish legal or financial output unverified. Bloomberg's report of internal scepticism about real-world coding fits the pattern: strong scores, uneven feel.

The most important precedent here is not a benchmark. Google is shipping a frontier model that is capability-gated to vetted cyber defenders. Its own model card states the unrestricted build shows a sharp uplift in dangerous cyber capability. That is why ASL-3 procedures were applied, and guardrails are removed only for trusted members of the Fairwind Program. The list includes CrowdStrike, Palo Alto Networks, Wiz, Mandiant, the US Air Force, Deutsche Telekom and 650+ others. The DeepMind blog also commits Google to STIX as its preferred threat-intelligence standard and 1,000+ Microsoft 365 hardening configurations. For GCC telecoms, banks and critical infrastructure the message cuts two ways. Top-end capability will increasingly be rationed through trust programmes that favour US-nexus vetting. Model behaviour is now negotiated per customer rather than uniform. Document your identity, logging and security posture now. Enter through your MSSP's membership where you have one.

How we would approach it: build against a route, not a model. Put an OpenAI-compatible gateway such as our FastLLM Proxy in front of every provider. Pin exact model versions per workload. Enforce a dollar budget per completed task, not per token. Keep a fallback to Opus 5.5 or Astra until Argon is GA on Vertex AI and you have re-run your own evaluation. Cap output tokens well below the 1M ceiling on first runs. Long agentic output compounds error as well as cost. A single runaway run now costs up to 15 times the old 64K limit.

# Policy sketch for an OpenAI-compatible gateway (FastLLM Proxy or equivalent)
routes:
  repo-migration:
    primary: $ARGON_ENDPOINT        # pin the exact model version once GA
    fallback: $OPUS55_ENDPOINT
    max_output_tokens: 200000       # do not start at 1M
    budget_usd_per_task: 1.00       # measured per completed task
    gate: "cost_per_task <= 1.21"   # beat Astra (max) or keep the fallback
  legal-summary:
    primary: $OPUS55_ENDPOINT
    max_output_tokens: 32000
    require_verification: true      # 15% hallucination is not zero
    budget_usd_per_task: 3.00

Most teams will get this wrong in one of two directions. They wait for general availability and lose three months of harness work. Or they rewrite prompts and agents around Argon's formatting before the API is public at all. The compounding advantage sits with organisations that treat the frontier as a rotating set of interchangeable backends. They invest now in evaluation, verification and gateway policy. In the GCC, where AI spend often runs through annual budget cycles and a single vendor relationship, that flexibility is also your negotiating position. Once cost per completed task is measured on your own code, you know exactly what a $0.74 migration run is worth against a $2.70 one. You know which model deserves the route.