Claude Opus 5.5 and GPT-6 Sol launch with big price cuts

2026-09-23 - ABikram Mondal

Price cuts arrive with the new models

Anthropic released Claude Opus 5.5 on September 22. The company said the model matches performance of its stronger Claude Fable 5.1 while cutting run costs by about 40 percent.

OpenAI followed hours later with GPT-6 Sol and GPT-6 Luna. Both sit at half the price of the prior GPT-5.6 versions. GPT-6 Luna lists at 0.10 dollars per million input tokens and 0.50 dollars per million output tokens according to reports from The AI News.

These moves mark the clearest price competition yet among the largest models. Earlier this month OpenAI had already placed GPT-6 Astra on Amazon Bedrock and other platforms at higher rates.

Reuters and Bloomberg both covered the releases within hours. Builders who track API bills noticed the shift first.

The timing lines up with reports of rising enterprise spend on frontier models. Lower prices change the math for teams running agents or long context work.

What Claude Opus 5.5 actually delivers

Anthropic positions Opus 5.5 as a cost-efficient option that keeps most of the capability of its top line. Internal tests showed it holds up on agent benchmarks and coding tasks that matter for professional use.

Pricing sits at 4 dollars per million input tokens and 20 dollars per million output tokens. That remains higher than the new OpenAI options but lower than the previous Opus generation.

Early tester notes from 9to5Google mention solid results on writing and frontend tasks. Some long-running agent jobs still show occasional drops in consistency compared with the heavier Fable 5.1.

The release avoids any claim of new benchmark records. It focuses instead on efficiency gains that let users run more queries for the same budget.

Companies already on Anthropic APIs can switch without code changes in most cases. The context window and tool use stay the same as the prior Opus line.

OpenAI expands the GPT-6 line

GPT-6 Sol and Luna extend the family that started with Astra earlier in September. Sol targets higher intelligence work while Luna prioritizes cost for volume tasks.

Sol carries higher rates than Luna but still undercuts the older Sol models by 50 percent. OpenAI highlighted gains on computer use and browser agent benchmarks.

Availability rolled out first to paid ChatGPT users and API customers. Enterprise plans on Azure and AWS Bedrock also gained access within the same window.

One internal OpenAI safety test showed reduced unintended outcomes on business scenarios compared with GPT-5.6. Exact numbers were not released publicly.

Developers testing the new models report faster iteration on agent workflows that previously hit cost walls. The price drop directly affects how many parallel agents a team can afford.

Who gains and who stays on the sidelines

Startups building agent products see the clearest upside. They can now afford longer context windows and more tool calls without blowing monthly budgets.

Teams already locked into fine-tuned older models may wait for migration testing. The new releases do not change fine-tuning options or data retention policies.

Heavy users of very large context or specialized coding agents still compare against GPT-6 Astra for raw capability. Astra remains the higher priced option in the OpenAI lineup.

Indian startups running on tight cloud spend get immediate relief. The lower rates make it practical to prototype multi-step agents that were marginal before.

Researchers tracking safety metrics note the price focus but see no new alignment details in these releases. The conversation on standards bodies continues separately.

Remaining gaps in the new models

Neither release claims to solve long-horizon planning or reliable multi-day agent runs. Testers still report drift on complex sequences that span dozens of steps.

Multimodal performance improves incrementally but does not reset benchmarks in image or video understanding. The focus stays on text and tool use.

Cost savings assume standard usage patterns. Heavy tool calling or very long outputs can still push bills higher than the headline rates suggest.

Some early comparisons show Claude Opus 5.5 edging GPT-6 Sol on certain writing tasks while Sol leads on structured agent benchmarks. Results vary by exact prompt and evaluation set.

No public data yet confirms how these models perform on the latest internal safety suites from other labs. Independent checks remain limited to what companies choose to share.

Practical next steps for teams

Start by running side-by-side tests on your highest volume agent workflows. Track token spend and success rate over the same tasks.

Update cost models to reflect the new rates before planning larger deployments. The difference compounds quickly at scale.

Monitor migration paths if you rely on specific tool formats or guardrails that differ between providers. Most SDKs handle the switch with minimal changes.

ABikram Mondal builds automation for exactly this kind of problem at https://abikrammondal.com/services/automation.

Watch for follow-up announcements on fine-tuning or enterprise features. These often arrive weeks after the initial model drops.

The short version. Anthropic and OpenAI cut frontier model prices sharply on September 22, making high-volume agent work far cheaper for teams that test the new options.

Sources

Reported from the sources above on 2026-09-23. Figures are as published at the time of writing. If something here has moved on, the linked source is the one to trust.

From the desk of ABikram Mondal

If you got here because you are actually thinking about putting models like this to work inside a real business, wired into the tools a team already uses, that is the work I do. I build for founders and small teams who want the thing to exist and work, not a deck about it.

Automation services · everything I do · talk to me