OpenAI and Anthropic Launch Cheaper Models Simultaneously

OpenAI has priced the GPT-6 Sol at $2 per million input tokens and $10 for output this Tuesday, half of what it charged for the GPT-5.6 Sol, which was $4 and $20. The smaller model, GPT-6 Luna, is now $0.10 and $0.50, compared to $0.20 and $1.20 in the previous generation. A company spokesperson told VentureBeat that these prices are permanent, not a launch promotional rate. On the same day, Anthropic released the Claude Opus 5.5 at $4 and $20, down from $5 and $25 for Opus 5. The most expensive models from both companies, GPT-6 Astra and Claude Fable 5.1, maintain a price of $10 and $50 per million tokens.

Both cuts were announced just hours apart and target the same expense line: caching. The cache read for Opus 5.5 dropped to $0.20 per million tokens, 60% lower than the $0.50 charged for Opus 5, according to Anthropic's release page. On OpenAI's side, the cache input also stands at $0.20 per million, with a writing fee of $2.50. For those running agents that review the same context with each step, this expense line tends to account for the largest portion of the monthly spend, not the output price that makes headlines.

What Changed in the Buyer’s Sheet

In comparison to Anthropic’s own peak, the Claude Fable 5.1 remains at $10 and $50 per million tokens, five times the Sol price at both ends. Google's Gemini 3.8 Flash operates at $0.75 and $3.75 during an introductory rate valid until December 31, rising to $1.50 and $7.50 on January 1, 2027. Any budget set for 2027 using today’s rates will be off by 100% for Google.

Anthropic claims that Opus 5.5 generates output over 30% faster than Opus 5 and costs 40% less under typical loads in standard settings. The 40% savings do not come from the list price, which decreased by 20%, but from the combination of cheaper caching with fewer reasoning tokens per task. This distinction impacts contract renegotiation: the list price cut is contractual, while the 40% depends on each client's usage profile and cannot be included in the clause.

The Benchmark Decided to Stop

In the Artificial Analysis Coding Agent Index, Sol scored 80 points, surpassing Claude Fable 5's 77.2. In OSWorld 2.0, Anthropic reports an 81.8% accuracy for Opus 5.5 in computer usage tasks. These are narrow margins, and OpenAI itself published an audit on July 8 indicating that about 30% of SWE-bench Pro tasks carry overly strict tests, incomplete statements, or misleading descriptions. Buying a model based on benchmark scores at this moment is akin to purchasing statistical noise.

There’s an opposing viewpoint worth mentioning: that the price per token has entered structural deflation and will continue to decline. The Gemini 3.8 Flash pricing contradicts this in January. Aggressive launch pricing is a tool for gaining load, and loads that have already migrated are costly to switch back. The $2 for Sol is valid today because OpenAI declared it permanent, not because the inference cost curve requires it.

Where the Account Changes

For Indian software factories, the effect is direct on margins. TCS, Infosys, and Cognizant sell application maintenance under fixed-price contracts priced per head, and have been embedding agent tools in these delivery centers. 50% cheaper inference improves the margin on the current contract and becomes a client-demanded discount in the next renewal cycle. This gain lasts through one renegotiation cycle.

In Brazil, companies like CI&T and shared services centers serving U.S. clients face the same equation with an added currency challenge: APIs are billed in dollars while many local contracts are in reais. A 50% price cut in dollars does not fully reach the P&L of those absorbing the mismatch.

In Germany and the rest of the European Union, the decision has ceased to be merely financial. Since August 2, 2026, the obligations of the AI Act apply to general-purpose models, and switching providers to capture half a dollar per million tokens requires revamping technical documentation and risk assessment. The cost of switching has risen just as the incentive to switch has increased.

The variable that no one publishes is how many tokens an agent consumes until completing a task. As suppliers compete on unit price and obscure consumption figures, the buyer remains without the single number that closes the account.

---

Editor's note (09/23/2026): The original headline of this article, "OpenAI and Anthropic Cut Top Prices on Same Day", has been corrected. The lower prices apply to new mid-tier models (GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5). The most expensive models from both companies, GPT-6 Astra and Claude Fable 5.1, kept their prices.