- Price Challenge: OpenAI founder and CEO Sam Altman says the company was willing to offer GPT-5.6 Sol at one-quarter of Claude Fable 5’s price.
- Published Rate: OpenAI’s published Sol rate remains $5 per million input tokens and $30 per million output tokens.
- Task Economics: Completed-task costs also depend on token efficiency, retries, integration, security, and how well each model fits the workload.
- Competitive Pressure: GLM Five-Two, V4-Pro, and Kimi K3 broaden customer options, while OpenAI must publish a lower rate before customers can test its challenge.
OpenAI founder and CEO Sam Altman is challenging Anthropic on July 14. He said OpenAI was willing to offer its GPT-5.6 Sol model at one-quarter of Anthropic’s Claude Fable 5 price, but OpenAI’s published rate card still does not include that deeper discount.
Because the deeper discount remains prospective, OpenAI’s current GPT-5.6 Sol price is still $5 per million input tokens and $30 per million output tokens. The AI model became generally available on July 9. Lower-priced Chinese alternatives like Kimi K3 and GLM-5.2 give developers and enterprise customers increasingly more leverage to negotiate or switch away from US AI labs..
Altman’s public comments also span uses and markets beyond model pricing. In a May 2026 discussion, he said Gen Z users sought planning and decision support from the service. Altman tied his current challenge to both price and efficiency:
“GPT-5.6 sol is half the price and ~twice as token efficient as fable in many cases for accomplishing the same task. happy to deliver at one-quarter of the price.”
Sam Altman, OpenAI founder and CEO (via Sam Altman on X)
What the Price Comparison Measures in Practice
Anthropic introduced Claude Fable 5’s usage-credit pricing with rates of $10 input and $50 output per million tokens. One million input tokens plus one million output tokens would produce a $35 raw Sol charge versus $60 at those Fable rates.
Even with those list prices, the comparison does not establish identical model capability or completed-task cost. Altman’s quarter-price language points to a larger discount, but so far these are only proclamations.
Beyond raw token charges, completing a job can require different prompt lengths, reasoning depth, tool calls, retries, and output sizes. Retrieval systems, security controls, monitoring, and orchestration add further costs between calls.
OpenAI’s model-cost evaluations attempts to capture part of that task-level difference. GPT-5.6 Sol currently cost roughly half as much as Claude Fable 5 for similar tasks on the Artificial Analysis Intelligence Index, a third-party model benchmark. It cost about one-quarter as much on the Agents’ Last Exam agent-task benchmark at medium reasoning.
Because benchmark results cannot establish the same advantage for every workload, application-level costs remain decisive. Perplexity CEO Aravind Srinivas summarized the constraint, stating: “The model alone is no longer the product.” A lower token rate loses value when repeated attempts or expensive integration consume the savings.
Cheaper Models Widen the Competitive Market
In the wider market, Altman’s OpenAI price war challenge landed as rivalry with Anthropic and Chinese model makers intensified. Zhipu AI’s GLM 5.2 is available at rates of $1.40 input and $4.40 output per million tokens.
GLM 5.2’s lower-cost open-weight positioning also lets organizations download and run the trained parameters. That option shifts costs toward hardware, energy, operations, security, and capacity planning.
DeepSeek’s V4-Pro offers standard rates of $0.44 per million input tokens and $0.87 per million output tokens after an earlier permanent price cut. Low rates do not establish capability parity, but adequate performance for a defined application gives customers negotiating leverage.
Moonshot AI’s open-weight Kimi K3 presents another tradeoff. Its benchmark results put it right behind Sol and Fable overall while ahead of several tested models on coding and general-agent benchmarks. Its 2.8-trillion-parameter architecture was built amid hardware constraints in China, but switching still requires tests of quality, application fit, security, and deployment cost.
Simon Koser, chief product officer at Tzafon also thinks “Cost has become a huge thing for some of these labs.” Separately, Microsoft is training its own sales staff to emphasize the efficiency and cost-effectiveness of its in-house models against OpenAI, Google, and Anthropic products. Providers now compete on list price, capability, required model usage, and serving expense; workload volume and operating ability determine whether a cheaper model reduces total costs.
OpenAI still has to offer a lower GPT-5.6 Sol rate before Altman’s quarter-price challenge becomes a customer option. Production workloads must then show whether fewer tokens, calls, and retries reduce completed-task costs enough to justify switching. Integration, security, and application fit remain part of that decision.
