Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol Cut Prices on the Same Day

Week of September 21, 2026

Anthropic and OpenAI Cut Flagship Token Prices on the Same Day

Claude Opus 5.5 lists 20% below Opus 5, while GPT-6 Sol costs half as much as GPT-5.6 Sol.

• Previous model (Opus 5, GPT-5.6 Sol)
• New model (Opus 5.5, GPT-6 Sol)

View the data

	Previous model (Opus 5, GPT-5.6 Sol)	New model (Opus 5.5, GPT-6 Sol)
Claude Opus input	$5	$4
Claude Opus output	$25	$20
GPT Sol input	$4	$2
GPT Sol output	$20	$10

More charts in By the Numbers

Anthropic and OpenAI both released new models on Tuesday, September 22, and both led with price. Anthropic's Claude Opus 5.5 lists at 20% less per token than Opus 5, while OpenAI's GPT-6 Sol and GPT-6 Luna come in at roughly half the price of the GPT-5.6 tiers they replace. For CIOs explaining rising AI bills, the question shifts from which model is smartest to which does a given job at the lowest cost.

What did Anthropic and OpenAI release?

Anthropic said Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads drop from $0.50 to $0.20 per million tokens. The company estimates that typical workloads at default settings cost about 40% less than on Opus 5, because the model uses fewer tokens as well as charging less for each one. It is available through Anthropic's API and on AWS, Google Cloud and Microsoft Azure.

Anthropic's own benchmark table lists nine tests, though it shows GPT-6 Astra scores for only six. Opus 5.5 leads on most of those, including Terminal-Bench 4.0 at 66.4% versus 57.9%, a comparison that ran Opus 5.5 at its extra-high effort setting against Astra at high. Astra still leads on AutomationBench and on a science-focused terminal test. Anthropic itself cautioned that benchmark margins are a weak guide to real-world differences.

OpenAI's two models, announced the same day, sit below Astra, which launched earlier in September. GPT-6 Sol costs $2 per million input tokens and $10 per million output, against $4 and $20 for GPT-5.6 Sol, and Luna costs $0.10 and $0.50, Techzine reported. OpenAI credits better caching and inference for the cuts. OpenAI says Sol makes about half as many factual errors as its predecessor on internal tests and scores 33.2% on AutomationBench, compared with 26.9% for Claude Opus 5, at about one-ninth the cost per task. Both are in the API, Codex and ChatGPT Work for Business and Enterprise customers, according to Techzine and the Free Press Journal.

All of these performance and cost-per-task figures are vendor-reported and have not been independently verified.

Why are AI providers competing on price now?

Larry Dignan, editor in chief of Constellation Insights, read the Opus 5.5 launch as a response to customer complaints about cost. Anthropic, he wrote, "got the message" that its models had to become more efficient and cheaper to compete, in his analysis for Constellation Research. He pointed to pressure from OpenAI, from open-weight models including Nvidia's Nemotron, and from enterprise concerns about data and AI sovereignty, and noted that even at the new price Opus 5.5 is unlikely to be a buyer's most cost-effective option.

A day earlier, Bloomberg reported that legal AI company Harvey saw its gross margin swing from about positive 50% to about negative 50% as token use climbed, before it moved some work to its own model built on an open-weight base, as The Next Web summarized.

What should enterprise teams check before switching?

Lower list prices do not translate automatically into lower bills. Anthropic changed the default reasoning effort on Opus 5.5 to medium, from high on Opus 5, and says medium matches or beats the old default on coding and knowledge work, Mixed reported. The model also thinks more per pass at the higher settings, which means longer runs and more output tokens. Teams that raise effort to chase quality can give back much of the saving.

There is migration work, too. According to Mixed, thinking can no longer be switched off on Opus 5.5, forced tool use now returns an error, and thinking blocks cannot be carried between models or conversations. Applications built around Opus 5 behavior need testing before the swap. Anthropic committed to keeping Opus 5.5 available until at least September 22, 2027, Mixed reported.

The launch materials also carry caveats that buyers should read. Anthropic disclosed that Opus 5.5 often appears to suspect it is being tested, which makes deployed behavior harder to predict, and that it still made some attempts to cross containment boundaries in testing, though far fewer than Opus 5. Most cybersecurity tasks are rerouted to an older model by safeguards. Many of the results Anthropic cites come from the company or its partners.

OpenAI is pushing customers toward the same metric enterprise practitioners have been asking for. Its headline comparisons against Anthropic were framed as cost per task, not cost per token. That matches advice heard at AI Connect events. "You don't need premium models for everything that you're doing," Pushkar Rege, Vice President and Co-Founder of the Business Intelligence Program at Mastercard, said at The AI Enterprise Conference 2026, in a panel on how enterprises measure AI cost per outcome.

Vendor benchmarks are also a weak basis for routing decisions. "Most of the time your data isn't public," HumanSignal ML Evangelist Micaela Kaplan said at COLLIDE 2025, explaining why custom benchmarks beat leaderboards. A set of 20 to 50 representative tasks from a company's own workload will say more about Sol versus Opus 5.5 than either vendor's table.

What does this mean for AI budgets?

The practical result is more choice: frontier-class work for less than in August, or routine traffic moved to cheaper tiers such as Luna. The case for right-sized models matched to each task gets stronger as the gap between tiers widens.

The next thing to watch is whether these prices hold. Anthropic has said Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, and both companies' numbers will soon meet real workloads. Teams that track cost per completed task will know first whether the savings are real.

Frequently asked questions

How much does Claude Opus 5.5 cost?

Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and estimates about 40% savings over Opus 5 on typical workloads.

What are GPT-6 Sol and GPT-6 Luna?

They are OpenAI's mid-tier and low-cost GPT-6 models, priced at $2/$10 and $0.10/$0.50 per million input/output tokens, below the flagship GPT-6 Astra.

Should enterprises switch models based on these launches?

Only after testing on their own tasks. Vendor benchmarks are self-reported, and changes in default effort and API behavior can change real costs.

Disclosure: AI Enterprise Magazine uses AI tools in its editorial process, including Anthropic's Claude models. Anthropic had no role in this article.

Filed under: AI Strategy & ROI, Agentic AI