Claude 5.5, GPT-6.1 Sol, Grok 4.7, and Gemini 4 Argon: API prices, cost per task, and which to use for what

Every figure here comes from the vendor's own page, read on October 14, 2026. Prices are standard-tier list prices in US dollars per million tokens. We did not benchmark these models ourselves. The cost-per-task and speed sections use third-party leaderboards and developer reports on X, and we label them as such. X posts are individual experiences: a sample, not a survey.

The table

Model	Input	Output	Cached input	Context	Max output
Claude Opus 5.5	$4	$20	$0.20	1M	128K
Claude Sonnet 5.5	$2	$10	$0.10	1M	128K
Claude Haiku 5.5, prompts up to 100K	$0.10	$0.50	$0.01	1M	128K
GPT-6 Sol	$2	$10	$0.20	1,050,000 (922,000 input max)	128K
GPT-6.1 Sol	$2	$10	$0.10	1,050,000 (922,000 input max)	128K
GPT-6 Luna	$0.10	$0.50	$0.01	1,050,000 (922,000 input max)	128K
Grok 4.7	$2	$6	$0.50	500K	No text output limit listed
Gemini 4 Argon, intro price	$2	$10	95% off input	1M unified input and output	Shares the 1M limit

Claude's context and max-output figures are from Anthropic's models overview. Its cache-read rate is 5% of input on Opus 5.5 and Sonnet 5.5. Grok 4.7's cached rate is from its release notes and Cursor's model table.

What each price sheet adds in the small print

Claude 5.5 family. Opus 5.5 fast mode is $8 input and $40 output, in Claude Code and the Claude Platform, per Anthropic's launch post. US-only inference is 1.1x, per claude.com/pricing. Haiku 5.5 moves to $0.50 input and $2.50 output above 100K-token prompts. Batch requests are 50% off. Opus 5.5 can no longer run with thinking off, and most cybersecurity tasks are routed to Opus 4.8 under its safeguards. Preserved thinking applies to Opus 5.5 for API accounts created on or after August 31, 2026.

GPT-6 Sol and Luna. Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output for the whole request. Cache writes are 1.25x input. Batch and Flex are half price, Fast mode is 2x, and regional processing adds 10%, per OpenAI's Sol and Luna model pages. OpenAI says the September 22 prices are half of GPT-5.6's promotional pricing (launch post). GPT-6.1 Sol (released September 29) keeps the $2/$10 price and halves cached input to $0.10. It has the same 272K surcharge, and its reasoning effort runs from low to max (model page). Its Ultrafast mode is 6x standard prices, set with service_tier: "ultrafast" (details in ChatGPT Pro's Ultrafast tier). OpenAI describes it as "near-Astra performance at a lower cost."

Grok 4.7. Above 200K prompt tokens, the API price is $4, $1 cached, and $12, per the release notes. The US regional endpoint adds 10%. The faster variant, Grok 4.7 Fast, is not on the public xAI API. It runs only in Cursor and Grok Build at 2x standard rates, per the developer guide. Reasoning effort goes up to xhigh.

Gemini 4 Argon. The $2/$10 price is introductory and ends November 30, 2026 (see what Argon's tokens really cost). After that, Google says $4 input and $20 output (launch post). The 1M limit counts input and output together, and Google points developers to a new continuation protocol. API access is through the Gemini API paid tier.

Reading the numbers

• On today's list prices, Sonnet 5.5, GPT-6 Sol, GPT-6.1 Sol, and Argon's intro rate all sit at $2 input and $10 output. Grok 4.7 matches the $2 input and lists $6 output.
• On December 1, Argon's posted price matches Opus 5.5's $4/$20, unless Google changes it.
• Haiku 5.5 and GPT-6 Luna share the same $0.10/$0.50 base rate. Their surcharges start at different points: 100K for Haiku, 272K for Luna.
• The context figures are not like-for-like. Argon's 1M is a combined input-and-output limit. OpenAI caps input at 922,000 of 1,050,000. Claude and Grok list a context window and a separate output limit.

What a task actually costs

Price per token is only half the cost. The other half is how many tokens a model spends to finish a job. Two independent leaderboards published cost per task this month. Their numbers differ because they use different harnesses and tasks, but they rank cost the same way.

Source	Model (setting, harness)	Score	Cost per task
Artificial Analysis Coding Agent Index, Oct 1	Claude Sonnet 5.5 (max) in Claude Code	68	$14.19
same	Gemini 4 Argon (high) in Antigravity CLI, promotional pricing	64	$5.84
same	GPT-6.1 Sol (xhigh) in Codex	63	$1.04
Arena agent leaderboard, Oct 2	GPT-6.1 Sol (Max)	#5 at the time	$0.56 median

Arena says GPT-6.1 Sol's cost was "65% lower cost than Claude Opus 5.5 (High), while landing within 2.59 pts" and "80% lower cost than Claude Sonnet 5.5 (Max), while landing within 1.29 pts." On Arena's agent leaderboard as of October 13, Claude Opus 5.5 (Max) ranks first at a $5.68 median cost per task, GPT-6.1 Sol (Max) ranks second at $0.65, and Gemini 4 Argon (High) ranks fifth at $0.84. Arena's ranges show the top models' error bars overlap, and the figures update continuously.

The pattern: Sonnet 5.5, GPT-6.1 Sol, and Argon have the same $2/$10 list price, yet their cost per task differs by up to 14x on the AA index. As one summary put it, "the 14x cost gap comes almost entirely from how many tokens each one spends per task." A small hidden-test run by one developer points the same way: "GPT-6.1 Sol: 985 tokens, $0.010 / Sonnet 5.5: 2.0k tokens, $0.020 / Opus 5.5: 2.5k tokens, $0.050 / All three: 40/40." The author cautions that these were short tasks, "so read it as a direction, not a leaderboard."

For Grok 4.7, the cost risk is context, not output. A $2 input price on a 400K-token repo is about $0.80 a call before caching. One builder put it this way: "Grok 4.7 is cheap right up until you fill the context." Above 200K tokens, the input rate also doubles to $4.

Speed and quality: what developers report

GPT-6.1 Sol is cheap but, users say, slow. This is the most consistent complaint in our sample. One account that tracks coding-agent calls posted a median of 18 tokens per second for 6.1 Sol against 89 for Opus 5.5. We could not verify that data, and no independent source we found publishes it. A Codex user with "137 billion tokens" reported "Sol gives me 20-30 tokens per second." OpenAI says its October 5 change made default speed ~50% faster through the ChatGPT subscription (post). That change was to the subscription, not the API. For API users, OpenAI's answer is Fast (2x) and Ultrafast (6x).

Opus 5.5 gets the "cleanest output" reports. Several developers who run several models on the same job describe a split. "Opus still writes the cleanest patches. Grok finds the weird bugs first. Sol and Astra are faster and argue about architecture," wrote one. A founder who ran the same task through both $200 plans reported "Codex 6.1 Sol took ~18hr to complete, while Opus 5.5 took just an hour to complete and have a much better results." That is one comparison on one task, not a benchmark. On Arena's Image-to-WebDev board, Opus 5.5 ranked first (Arena).

Argon answers fewer factual questions wrong. On AA-Omniscience, Artificial Analysis reports a 15% hallucination rate for Argon, "the lowest of any model scoring 45+ on the Intelligence Index," versus 51% for GPT-6 Astra (max) and 54% for GPT-6.1 Sol (max). Its accuracy was lower, at 50% against Astra's 63%.

Grok 4.7 is the most divisive. Reports range from "finds the weird bugs first" to a benchmark account's "the output is still not usable for real frontend work". The themes are in Grok 4.7 in Cursor: what builders say.

Which to use for what

This is a summary of the evidence above, not a recommendation. Run your own workload.

• High-volume agent coding on a budget: GPT-6.1 Sol has the lowest measured cost per task of the models listed on both leaderboards. The trade-off developers report is speed.
• When patch quality matters more than cost: Opus 5.5 ranks first on Arena's agent board and, per Arena's own post, its Image-to-WebDev board, at about 9x Sol's median cost per task on Arena's October 13 figures.
• Top score on the AA index regardless of cost: Sonnet 5.5 (max) scored 68, at $14.19 a task.
• Factual or business-workflow tasks where "I don't know" beats a guess: Argon, on Artificial Analysis's hallucination numbers. Budget for the price doubling on December 1 (Argon's real token price).
• Large context on a tight budget, or bug hunting as a second opinion: Grok 4.7 has the lowest output price here ($6). Keep prompts under 200K tokens to avoid the $4 rate.
• Classification, extraction, and routing: Haiku 5.5 or GPT-6 Luna at $0.10/$0.50.

If you are deciding between paying per token and a subscription, see ChatGPT Pro's three prices, and Claude Max from $100. Subscription plans that include these models are on AI Model & Plan Prices.