Short answer: for most teams, pick the model whose harness you already live in — the price won't break the tie, because there isn't one. As of October 13, 2026, Gemini 4 Argon, GPT-6.1 Sol, and Claude Sonnet 5.5 all cost $2 per million input tokens and $10 per million output. When three frontier models carry the same sticker, the sticker stops being a decision input. Choose on the three things it hides: the harness your team works in, your cache-read economics, and — for Argon specifically — the introductory-rate clock. Settle it on tokens-per-task from your own repo, not the launch-day charts.
Here's the whole decision in one screen:
- Price is a three-way tie. All three are $2/$10. Stop comparing the rate card; it's the same card — with one asterisk below.
- The asterisk is Argon's. Google lists a standard rate of $4/$20 after an unpublished introductory period. Budget Argon at double, and the promo becomes upside instead of a surprise.
- The harness is the real switch. Sonnet 5.5 → Claude Code. Sol → Codex. Argon → Antigravity and Google AI Studio. Whichever your team already works in moves velocity more than the model does.
- The real bill is below the sticker. A coding agent re-sends the same context every turn, so cache-read pricing and tokens-per-task drive your cost far more than the matching $2/$10.
- Don't trust a single benchmark. Only Anthropic published specific agentic-coding numbers; Argon's and Sol's circulating figures come from third parties and conflict. Test on your codebase.
Why the identical price changes the question#
For two years, "which model?" was mostly a cost question, because frontier models were priced far apart and you traded quality for dollars. That trade is gone for this tier. Anthropic shipped Sonnet 5.5 at $2/$10 on September 28; OpenAI matched it with GPT-6.1 Sol a day later; and on October 13 Google opened Gemini 4 Argon — a model it had kept gated since September 30 — at the same number. Three of the strongest agentic-coding models in the field, all on the exact same rate.
When the rate is identical, the interesting differences are the ones the rate card hides. There are three that actually decide it.
1. The harness moves your velocity more than the model#
Each model ships inside a coding agent, and that agent — not the weights — is what your engineers touch all day. Sonnet 5.5 runs natively in Claude Code and is also on the Anthropic API plus AWS Bedrock, Google Vertex, and Azure Foundry with a zero-data-retention option. GPT-6.1 Sol runs natively in Codex and ChatGPT Work. Gemini 4 Argon runs in Google's Antigravity editor (Ultra tier) and AI Studio, reachable through the Gemini API for anything you build yourself.
How each harness edits files, runs your tests, reviews a diff, and asks permission before a destructive command will change throughput more than a two-point benchmark gap ever will. The practical read: if your team already lives in one, that's a strong default — the switching cost of retraining habits and rebuilding config usually swamps the model delta. If you're greenfield, pick the ecosystem whose compliance and cloud posture you already have: Sonnet's multi-cloud reach, Sol's OpenAI standardization, or Argon's Google Cloud footprint.
2. The real cost lever is below the sticker — and Argon has a clock#
A coding agent doesn't send one prompt. It loops, re-sending the same file tree, system prompt, and tool definitions on every turn — tens of thousands of repeated input tokens per task. That's why cached-read pricing and tokens-per-task swing your bill far more than the base rate that happens to match. All three discount cached input heavily (Argon's cached reads are 95% off; Anthropic lists Sonnet's cache reads at roughly a fifth of the input rate), and Anthropic's pitch for Sonnet 5.5 is explicitly token efficiency — "up to 30% less per task", which is a lower bill at the same rate, not a lower rate. Your true cost is (tokens the model burns) × (rate), discounted by cache hits — and that product can differ by multiples between two models sharing one sticker. It's the same logic behind why cache-read pricing quietly dominates an agent's bill and routing each request to the cheapest capable model.
Argon carries the one pricing asterisk worth circling: its $2/$10 is introductory. Google's footnote lists a standard rate of $4/$20 once that period ends, and no end date has been published. If you pick Argon, model it at $4/$20 from day one. A promo you planned around is upside; a promo that expires on a date you didn't know is a budget hole. OpenAI's counter-lever, meanwhile, is speed: an Ultrafast tier that clocks up to ~300 tokens/sec in Codex at a premium rate. If a human is waiting on the generation, speed may be worth it; if your invoice is the bottleneck, token efficiency wins.
3. Don't let a launch benchmark pick for you#
This is where the three genuinely differ in what they've disclosed, not necessarily in what they can do. Anthropic published specific agentic-coding benchmarks for Sonnet 5.5 — Terminal-Bench 4.0 at 70.6%, CursorBench 4.0 at 55.5%, OSWorld 2.1 at 80.1%. Google positions Argon as a frontier model for complex, multi-step work across coding, finance, law, cybersecurity, and science, and one third-party index ranked it first at launch — but the per-benchmark numbers on aggregator sites already disagree with one another, the usual sign the ranking isn't settled. Sol's coding figures are mostly third-party and on mismatched benchmark versions.
A model's benchmark page tells you what it scored on someone else's tasks. Your repo is the only benchmark that bills you.
Critically, none of the three cleanly reported SWE-bench Verified head-to-head at these launches, so any "X beats Y on SWE-bench" chart you see is not from the launch data. The reliable move is the same discipline as reading a coding benchmark critically before trusting the headline: a public score is a starting hypothesis, your repository is the experiment.
How to actually pick, this week#
- Take 10 representative tickets from your real backlog — the mix you actually ship, not toy problems.
- Run each through all three native harnesses: Claude Code (Sonnet 5.5), Codex (Sol), and Antigravity or AI Studio (Argon).
- Record three numbers per model: did it pass, total tokens per task, and cache-read hit rate.
- Compute true cost as
tokens-per-task × rate, adjusted for cache hits — and use $4/$20 for Argon, not the promo. This is where the identical sticker fractures into a real difference. - Weigh the harness ergonomics your team will live with every day — the permission model, the diff review, the test loop.
Three models, one price. The launch charts want you to pick a winner; your backlog will actually pick one. For the news peg behind this — why three labs landed on the same number, and what it says about where the model layer is headed — see this morning's Founder's Wire.



