Analysis

Opus 5.5 at $4 and GPT-6.1 Sol at $2: What Cheaper Flagship Pricing Does to Unit Economics

Anthropic cut its flagship rate by 20% while OpenAI priced GPT-6.1 Sol at half of Opus 5.5. List price is only the start of what a task costs.

Tobias Reyes

AI Industry & Policy Analyst

Published 7 min read
Organized office desk with pricing formula document, laptop, and smartphone calculator.
In this story 9 sections

Claude Opus 5.5 lists at $4 per million input tokens and $20 per million output tokens, while OpenAI's GPT-6.1 Sol lists at $2 and $10. Both sit well below their labs' premium tiers. List price is only part of the bill, because tokens used per task, caching and long-context surcharges decide what buyers actually pay.

This analysis is for engineering leaders, finance teams and product owners who buy frontier model access and need to compare two launches that arrived a week apart. It covers the posted prices, a worked cost example, the pricing rules that change the math, and how to test claims on your own workload. Emergent Wire starts from unit economics, and the launch coverage so far has focused on benchmarks and not on what a finished task costs.

What Are the Posted Prices for Opus 5.5 and GPT-6.1 Sol?

A detailed view of a blue lit computer server rack in a data center showcasing technology and hardware.

Anthropic launched Claude Opus 5.5 on September 22, 2026. Launch coverage lists API pricing of $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, and a cache-read price that fell from $0.50 to $0.20 per million tokens, according to Digg's report on the launch. Anthropic says the model costs 40% less to run than Opus 5 and performs at the level of Claude Fable 5.1 on most work.

OpenAI released GPT-6.1 Sol on September 29, 2026, at its developer event. DataCamp's guide lists it at $2.00 input and $10.00 output per million tokens, with cached input at $0.10 and cache writes at $2.50, per its GPT-6.1 Sol breakdown.

What Are the Posted Prices for Opus 5.5 and GPT-6.1 Sol?
ModelInput per millionOutput per millionCached input per million
Claude Opus 5.5$4$20$0.20
Claude Opus 5$5$25$0.50
Claude Fable 5.1$10$50$0.25
GPT-6.1 Sol$2$10$0.10
GPT-6 Astra$10$50Not listed

The Fable 5.1 figures come from Stokastic's launch coverage, and the Astra figures from DataCamp. Prices come from secondary sources, so confirm them on each vendor's pricing page before you budget.

How Does List Price Turn Into Cost Per Task?

Multiply tokens by rate. Tokens, not the sticker price, drive the bill. As an illustration, take an agent task that reads 150,000 input tokens and writes 15,000 output tokens, with no caching.

How Does List Price Turn Into Cost Per Task?
ModelCost per taskCost per 10,000 tasks
GPT-6.1 Sol$0.45$4,500
Claude Opus 5.5$0.90$9,000
Claude Opus 5$1.13$11,250
GPT-6 Astra or Fable 5.1$2.25$22,500

This is a worked example, not a measurement. On that workload, Sol costs half as much as Opus 5.5, and Opus 5.5 costs about 20% less than Opus 5. Both are far below the $10 and $50 tier. The ranking only holds if each model uses the same number of tokens, which is rarely true.

Why Do Real Costs Differ From the Math?

Business professionals engaging in a collaborative meeting with charts and documents.

Agents do not use a fixed number of tokens. A model that explores more, retries more or reasons longer burns more tokens per task. Reasoning tokens bill at the output rate on Sol, according to DataCamp, so extra thinking is billed at the highest price.

Independent cost-per-task data shows the spread. In Artificial Analysis's coding index, as summarized by CellCog, Claude Code with Opus 5.5 cost about $11.42 per task, while Codex with GPT-6.1 Sol cost about $1.50. Those are real agent runs at maximum effort, and they differ by a factor of seven, far more than the 2x list-price gap. We cover that data in our guide to the Coding Agent Index v2.0.

The lesson is that list price sets a floor and behavior sets the bill. A cheap model that needs many steps can cost more than an expensive one that finishes quickly.

Which Pricing Rules Change the Math?

Three rules matter most. The first is the long-context surcharge. For prompts above 272,000 input tokens, Sol charges double on input and cache rates and 1.5 times on output for the whole request. At that size Sol costs $4 input and $15 output, so its input rate equals Opus 5.5's, while its output is still cheaper.

The second is caching. Cached input is $0.10 on Sol and $0.20 on Opus 5.5. A workload that reuses a large prompt, such as a codebase or a policy manual, pays mostly cache rates after the first request. Cache design can cut a bill more than a model switch.

The third is promotions and modes. DataCamp reports promotional pricing on Sol through at least November 21, 2026, and a fast mode priced at twice the standard rate. Check which mode your product actually uses, because a faster tier can double the rate.

Is a Cheaper Flagship Also a Weaker One?

A pink piggy bank symbolizes savings and budgeting on a simple brown backdrop.

Not necessarily, but the evidence is thin. Anthropic reports 66.4% for Opus 5.5 on Terminal-Bench 4.0 and 54.4% on FrontierCode v1.1, per Stokastic. OpenAI says Sol matches GPT-6 Astra on DeepSWE v1.1 and scores 2.2 points above Opus 5.5 on AutomationBench at roughly a third of the cost, per DataCamp.

These are vendor-reported figures. DataCamp says nearly all of its numbers are OpenAI's own, with no independent reproduction published. Evaluation setups also differ. Anthropic's FrontierCode score of 54.4% sits below the 56.0% in the independent index, so the same benchmark can produce different results under different harnesses.

The premium tier still exists for a reason. OpenAI's own data shows Astra ahead of Sol on several hard tasks, such as TroubleshootingBench at 63.46% against 47.96%. If your work is a hard, high-stakes task where one error costs more than the token bill, the premium may pay for itself.

What Does the Price Cut Say About the Business?

It says capability is moving down the price curve quickly. Anthropic's claim that Opus 5.5 performs at Fable 5.1's level on most work, at 60% lower rates, means last month's premium tier is becoming this month's mid-tier. OpenAI is doing the same with Sol against Astra.

For labs, this has a cost. Lower prices squeeze margins unless the cost of serving falls as fast. For buyers, it means annual contracts priced at last quarter's rates may be too high. We looked at the drivers in our explainer on why AI model prices keep falling and in our guide to inference cost per million tokens.

Distribution matters too. When labs sell through several clouds, buyers can compare prices more easily, as we explained in our piece on OpenAI models arriving on Google Cloud. More comparison points push prices down faster.

Who Actually Pays for Frontier Access?

The buyer mix explains the pricing. Heavy API users, such as coding tools and support platforms, run millions of tasks and feel every cent per million tokens. Occasional users pay through subscriptions and rarely see token prices. Labs set API rates for the first group and use plan limits for the second.

That split is why usage limits matter as much as price. For a subscriber, a higher limit works like a price cut, so check plan limits as well as API rates.

For enterprise buyers, the decision is closer to procurement than to a model choice. The team that owns the budget should know which workloads need the premium tier and which can run on the cheaper one. A routing rule, such as sending routine tasks to Sol or Opus 5.5 and hard ones to a premium model, often beats using a single model for everything.

How Should Buyers Test the Claims?

Overhead view of a team analyzing graphs and charts using laptops at a meeting.

Measure cost per finished task on your own workload. At Emergent Wire, we treat list price as the start of a calculation, not the end. Follow these steps:

  1. Collect 30 real tasks, with an automatic pass or fail check for each.
  2. Run each task three times on each candidate model.
  3. Record input tokens, output tokens, cached tokens and reasoning tokens.
  4. Calculate cost per passed task, not cost per attempt.
  5. Repeat at your long-context sizes, since surcharges apply above set limits.

Include human review time in the calculation. A model that passes 90% of tasks at $0.45 may cost more than one that passes 95% at $0.90, if each failure needs ten minutes of engineer time. Hourly engineer cost often outweighs token cost.

Also lock in contract flexibility. Ask providers whether prices can drop mid-term and whether committed spend can move across models. A short term with a price-protection clause beats a long term at today's rate.

Conclusion

Opus 5.5 at $4 and $20 and GPT-6.1 Sol at $2 and $10 show that frontier capability is getting cheaper by the month, but list price hides the real driver, which is tokens per finished task. Emergent Wire's editors compared the posted rates with independent cost-per-task data and found a gap far larger than the sticker difference.

Your next step is a small paid trial. Pick 30 tasks, run both models three times, and decide on cost per passed task. Revisit the decision each quarter, because prices are moving faster than most contracts.

How much does Claude Opus 5.5 cost per million tokens?
Reports on Anthropic's launch put Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads fell from $0.50 to $0.20 per million tokens. Anthropic launched the model on September 22, 2026.
How much does GPT-6.1 Sol cost?
DataCamp lists GPT-6.1 Sol at $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.10. Prompts above 272,000 input tokens cost double on input and 1.5 times on output for the whole request. OpenAI released the model on September 29, 2026.
Is the cheaper model always the cheaper choice for a task?
No. A task's cost depends on how many tokens the model uses, not just the per-token price. A model with a lower rate that needs more steps, retries or reasoning tokens can cost more per finished task. Compare cost per solved task on your own workload, not list prices.
Should companies switch to the cheaper flagship models?
Not automatically. Benchmarks for both models come mostly from the vendors, and outside reproductions are limited. Run a small trial on your own tasks, measure pass rate and cost per finished task, and keep a premium model available for the hardest work where a lower error rate saves more than it costs.

Written by

Tobias Reyes

AI Industry & Policy Analyst

12+ years experience

Tobias Reyes spent ten years as a technology strategy consultant, most of that time advising Fortune 500 clients on enterprise software adoption before AI became the whole conversation. He joined Emergent Wire to cover the industry from the incentive layer up: who actually pays for frontier model access, how compute costs shape product decisions, and where policy is likely to reshape the competitive landscape next. He is allergic to roadmap-slide analysis and prefers to start from unit economics and public filings whenever they exist.

Credentials

  • Former technology strategy consultant (10 years, enterprise clients)
  • MBA, focus on technology and operations strategy
  • Regular commentator on AI industry economics and policy