TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
OpenAI and Anthropic cut prices across several model families between June and September 2026, with some listed rates falling by as much as 80%. The timeline points to cheaper serving costs, particularly for cached prompts, and pressure from rival models, but the companies have not disclosed a single cause for the broader price reductions.
OpenAI and Anthropic lowered prices for several large language models between June and September 2026, with some rates dropping by as much as 80% from earlier levels. The sequence has made comparable models cheaper to use, but the available reporting points to several possible forces—not one confirmed explanation—including lower serving costs, competition and the growing importance of cached-token pricing.
OpenAI’s June 26 release of GPT-5.6 Sol kept the listed price at $5 per million input tokens, $30 per million output tokens and $0.50 per million cached-read tokens. Its smaller Luna model launched at $1, $6 and $0.10, respectively. On July 30, OpenAI cut Luna’s rates by 80%, to $0.20 input, $1.20 output and $0.02 cached reads. The company later reduced Sol’s rates for three months, then introduced GPT-6 Sol in September at $2 input and $10 output per million tokens.
Anthropic also made changes across successive releases. Claude Sonnet 5 launched June 30 at $2 input, $10 output and $0.20 cached reads per million tokens; the company initially described that price as temporary, then made it permanent on August 10. Claude Opus 5.5, announced September 22, was priced at $4 input and $20 output, with cached reads at $0.20. Anthropic explicitly said Opus 5.5 requires less compute to serve than Opus 5 and that its price reflects this.
Some of the largest effective savings are tied to cache reads, not just headline input and output rates. The source report says cached tokens account for 96%–99% of token expenditure in the coding-agent workloads it discusses, and that cache reads are often hit 90%–100% of the time. It estimates that a 75% cut in cache-read pricing can lower total model-use costs by about 70%–75% for those workloads. These are workload-specific estimates, not a universal discount for every use case.
Cheaper Tokens, More Agent Use
Lower per-token prices can reduce the cost of using AI systems for coding agents and other multi-step tasks, which repeatedly send conversation history back to a model. For those applications, cached-read rates can matter as much as—or more than—the standard input and output prices. Price changes may also affect use through subscription-based tools such as Codex and Claude Code, according to the source report.
Cheaper usage does not automatically mean lower total spending. The report cites Vercel AI Gateway data showing cost per token at 43.3% of its April 2026 level, while aggregate spending increased. The report attributes that pattern to model improvements driving greater use, but the cited data alone does not establish why total spend rose. For customers, the practical question is whether the lower unit price offsets higher usage and whether a particular model remains suitable for the task.
Products Worth Considering
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
A Rapid Series of Model Repricing
The price changes followed a compressed run of releases. OpenAI announced GPT-5.6 models in June, then cut Luna’s price in July. Anthropic made Sonnet 5’s introductory price permanent in August and released Fable 5.1 in September with a cached-read rate of $0.25 per million tokens, down from $1 for Fable 5. OpenAI’s GPT-6 Astra followed at $10 input and $50 output per million tokens, with a $1 cache-read rate.
On September 22, OpenAI priced GPT-6 Sol at half the input and output rates of the temporarily discounted GPT-5.6 Sol, and priced GPT-6 Luna at half the previous Luna rates. A week later, OpenAI replaced GPT-6 Sol with GPT-6.1 Sol and halved its cached-read price to $0.10 per million tokens. The source author says this made the family cumulatively 75% cheaper, comparing the relevant rates across releases.
The source report describes the timing as a tit-for-tat pattern between OpenAI and Anthropic, but that is an interpretation of the release sequence, not a company-confirmed account of decision-making. It also points to substitute models, including open-weight options such as DeepSeek, as sources of competitive pressure. The author argues that users can switch models relatively easily through agent tools or API gateways, making price and performance comparisons more immediate.
“Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.”
— Anthropic, describing Claude Opus 5.5
What Is Driving the Full Price Shift
No single cause is confirmed for the broad series of cuts. Anthropic directly linked Opus 5.5’s pricing to lower serving compute, but the source material does not provide equivalent explanations for every OpenAI and Anthropic price change. The proposed role of rival pricing is plausible from the timing, yet neither company is quoted confirming that it cut rates in response to the other.
It is also unclear how broadly the stated savings apply. Actual bills depend on the model, workload, token mix, cache-hit rate, provider infrastructure and usage volume. The source report says output quality did not appear to decline alongside the cuts, but that is the author’s assessment; the material does not establish a comprehensive, independent comparison across tasks or users. The excerpt also does not give enough detail to verify all cross-model performance comparisons.
Watch Rates and Real-World Costs
Customers and developers will need to compare current rate cards and total task costs, including cached reads, rather than relying only on launch prices or input-token rates. OpenAI’s Sol discount was described as lasting three months, so its status after that period is a point to check. Further model releases or rate changes could alter the comparisons quickly.
The next useful evidence would be clearer company explanations for pricing decisions and independently measured costs across representative workloads. Until then, the confirmed picture is a marked reduction in listed prices across multiple releases, with lower serving costs explicitly cited for one Anthropic model and competitive pressure remaining a reasonable—but unverified—part of the explanation.
Key Questions
Which prices fell by as much as 80%?
OpenAI cut GPT-5.6 Luna’s listed input, output and cached-read rates by 80% on July 30, 2026, according to the source report.
Did OpenAI and Anthropic confirm they were responding to each other?
No. The source report interprets the close timing of the changes as a tit-for-tat pattern, but neither company is quoted confirming that explanation.
Why can cache-read pricing make such a difference?
Coding agents often resend prior conversation context, which can be served from a cache. The report estimates cached reads make up 96%–99% of token expenditure in the agent workloads it discusses, so lower cache rates can sharply reduce costs for those workloads.
Do lower token prices mean customers will spend less overall?
Not necessarily. The report cites Vercel data showing a lower cost per token alongside higher aggregate spending. More usage can offset lower unit prices, and total costs vary by workload and provider.
Is model quality confirmed to be unchanged after the price cuts?
No comprehensive independent finding is provided in the source material. The author says quality did not appear to be sacrificed, but that remains an assessment rather than a verified conclusion across all tasks.
Source: hn
