Which $2/$10 model is best: Gemini 4 Argon, Claude Sonnet 5.5 or GPT-6.1 Sol?

On the independent Vals Index, Gemini 4 Argon leads at 68.90% versus 67.04% for Claude Sonnet 5.5, with GPT-6.1 Sol not yet in the top group; on Artificial Analysis, Sonnet 5.5 scores 56 versus 52 for GPT-6.1 Sol, with no Argon score yet. Sonnet 5.5 wins terminal agents (Terminal-Bench 4.0 64–70.6%), GPT-6.1 Sol wins cost per task ($0.72 per AA index task), and Argon wins breadth and output length — but Argon is not publicly available as of October 1, 2026.

Is Gemini 4 Argon's $2/$10 price permanent?

No. Google calls it an introductory price and says Argon will move to $4 per million input tokens and $20 per million output tokens, without giving a date. Claude Sonnet 5.5's $2/$10 is a standard list price (Anthropic cancelled a planned Sonnet increase in August 2026), and GPT-6.1 Sol's $2/$10 is OpenAI's standard rate. Cached input differs: Argon $0.10 (95% off), GPT-6.1 Sol $0.10, Sonnet 5.5 $0.20.

Which one can I actually use right now?

Claude Sonnet 5.5 (released September 28, 2026) is on the Anthropic API, AWS, Google Cloud and Azure. GPT-6.1 Sol (September 29, 2026) is on the OpenAI API and in ChatGPT Work and Codex on paid plans, but not in regular ChatGPT yet. Gemini 4 Argon (September 30, 2026) is restricted to Google's Fairwind Program cyber defenders; paid API and Google AI Ultra access are promised 'as soon as possible' with no date.

Which has the biggest output limit?

Gemini 4 Argon: Google says 1M output tokens in a single trajectory, up from 64K on earlier Gemini models (Vals evaluated at 262K). Claude Sonnet 5.5 and GPT-6.1 Sol are both capped at 128K output tokens. All three have ~1M-token context windows; GPT-6.1 Sol reprices prompts above 272K tokens.

Quick Answer

Gemini 4 Argon vs Sonnet 5.5 vs GPT-6.1 Sol: $2/$10 Tier

Published: October 1, 2026

The short answer

Three frontier-adjacent models now list at exactly $2 per million input tokens and $10 per million output: Claude Sonnet 5.5 (September 28, 2026), GPT-6.1 Sol (September 29) and Gemini 4 Argon (September 30, introductory). Argon leads the Vals Index (68.90% vs 67.04% for Sonnet 5.5) and has an 8x larger output limit, but is gated to cyber defenders and will rise to $4/$20. Sonnet 5.5 is the best public terminal-coding model in the tier; GPT-6.1 Sol is the cheapest per task and has the cheapest cache. Facts verified October 1, 2026.

Side by side

	Gemini 4 Argon	Claude Sonnet 5.5	GPT-6.1 Sol
Released	Sep 30, 2026	Sep 28, 2026	Sep 29, 2026
Input / output	$2 / $10 intro → $4 / $20	$2 / $10 (standard)	$2 / $10 (standard)
Cached input	$0.10	$0.20	$0.10
Context	1M	1M	1.05M (>272K reprices)
Max output	1M (Google)	128K	128K
Availability	Fairwind defenders only	API, AWS, GCP, Azure	API, ChatGPT Work, Codex
Vals Index	68.90% (#1 of 41), $15.68/test	67.04% (#2), $21.34/test	Not in top 4
AA Intelligence Index	Not scored	56	52
AA cost per index task	—	$7.60 (max effort)	$0.72
Terminal-Bench 4.0	57.4% (Google) / 57.58% (Vals #5)	64% (AA) / 70.6% (Anthropic)	— (GPT-6 Astra 60%)
DeepSWE v1.1	77.9%	—	matches Astra (74.1%) per OpenAI
Vals Vibe Code Bench	91.91% (#2)	92.39% (#1)	—
Vals Code Migration	68.17% (#2)	69.83% (#1)	—
Vals CyberBench	77.86% (#2)	—	GPT-6 Sol #1 (+0.12)
Output speed (AA)	—	~138 tok/s	~67 tok/s
Gray Swan prompt-injection ASR	0.7%	— (Opus 5.5: 1.0%)	— (GPT-6 Sol: 27%)

Vals scores as read October 1, 2026; AA scores from Artificial Analysis Intelligence Index v4.3.2 as of September 30, 2026.

Same price, three different bets

Gemini 4 Argon is a frontier model sold at mid-tier pricing — for now. Google benchmarks it against GPT-6 Astra ($10/$50) and Opus 5.5 ($4/$20), not against Sonnet or Sol, and wins 13 of 18 rows. Its $0.10 cache and 1M output limit make it the obvious choice for long single-trajectory jobs like codebase migrations, if you can get it and if the intro price holds. Vals’ per-test cost ($15.68) is 27% below Sonnet 5.5’s, but long agentic tests ran to $57.82 (Code Migration) and $193.78 (CUA-bench) because the model uses its output headroom.

Claude Sonnet 5.5 is the best model in the tier you can deploy today for agentic coding: 64% on Terminal-Bench 4.0 on AA (Sonnet 5 was 14%), #1 on Vals Vibe Code Bench and Code Migration, and 138 tok/s output. The cost is tokens: at max effort it emits ~193K output tokens per AA task — $7.60 per task, more than 10x GPT-6.1 Sol. At high effort AA says it sits “very narrowly behind GPT-6 Sol at effectively the same cost per task.”

GPT-6.1 Sol is the volume play. OpenAI says it matches Astra on DeepSWE v1.1 at one-fifth the cost, and AA measured $0.72 per index task. Cached input at $0.10 halves Sonnet’s for agents that resend big system prompts. The weakness is prompt-injection resistance — Google’s chart shows GPT-6 Sol at a 27% attack success rate versus 0.7% for Argon — and OpenAI replaced GPT-6 Sol after seven days, so budget for churn.

The cache maths

For an agent re-sending a 25K-token system prompt and tool list every turn: Argon and GPT-6.1 Sol cost $0.0025 per turn on cache hits; Sonnet 5.5 costs $0.005. At 10,000 turns a day that is $25 vs $50 — trivial next to output spend, decisive at fleet scale.

Which to use

Terminal-heavy coding agent you watch: Claude Sonnet 5.5.
Unattended agent fleet billed by the token: GPT-6.1 Sol.
Long-horizon migrations, legal/finance research, or you are a Fairwind partner: Gemini 4 Argon — and lock in the intro price while it lasts.
Need prompt-injection resistance for tool-using agents: Argon (0.7%) or an Anthropic model (1.0%); avoid Sol-tier OpenAI models for untrusted-content agents until OpenAI publishes a fix.

Last verified: October 1, 2026. Prices from vendor pages; scores from Vals AI and Artificial Analysis.

The short answer

Side by side

Same price, three different bets

The cache maths

Which to use

Sources