Direct answer
API users save money; ChatGPT subscriptions did not get cheaper
If you pay for OpenAI API usage, short-context Standard Luna is now $0.20 per million input tokens and $1.20 per million output tokens; Terra is $2 input and $12 output. OpenAI cut those rates by 80% and 20% on July 30. Ordinary ChatGPT subscription prices did not fall.
Keep Standard when it already meets the deadline. Evaluate Luna or Terra on a fixed workload when token volume matters, and reserve Fast mode for jobs where shorter response time is worth a 2x token-rate premium. OpenAI says Sol Fast can run up to 2.5 times faster; that is a provider claim, not a TPTK test.
Decision first
Keep Standard, route lower-risk work, or pay for speed
Your production path already meets its deadline
Do not pay a latency premium when the current queue, Batch, Flex, or Standard service already meets the business response-time requirement.
Volume makes token cost material
Re-test repeatable, lower-risk, or well-specified stages where the lower rates could reduce cost without increasing retries, human review, or failure expense.
Latency has a measured dollar value
Pay the premium only when faster Sol responses protect a customer interaction, operator time, deadline, or other outcome worth more than the added API cost.
What changed July 30
Two price cuts and one service-tier rename
Short-context Standard input moved from $1 to $0.20 per million tokens; output moved from $6 to $1.20.
Short-context Standard input moved from $2.50 to $2; output moved from $15 to $12.
Sol remains $5 input and $30 output per million short-context tokens. Fast mode is a separate premium tier.
OpenAI also says ChatGPT and Codex subscription prices and quota budgets did not change. Terra and Luna now consume fewer credits in paid ChatGPT Work and Codex subscriptions. That distinction matters: this is an API price and credit-consumption change, not a subscription-price reduction.
Current direct OpenAI rates
Use the service tier and context band that matches the request
The cards below show OpenAI’s current per-million-token list rates checked August 3. They are planning inputs, not a quote. Regional processing carries a 10% uplift for eligible models released on or after March 5, 2026, and Amazon Bedrock billing may differ from direct OpenAI pricing.
Boundary: Batch and Flex rates are lower than Standard in OpenAI’s current table, while Fast rates are higher. Long-context pricing changes the input/output mix. Tool fees, cache behavior, reasoning/output length, retries, taxes, contracts, and cloud-provider rates can change the invoice.
Reproducible cost check
A 100M-input, 20M-output workload drops from $220 to $44 on Luna
This example applies the launch-era and July 30 short-context Standard token rates to one disclosed hypothetical workload: 100 million uncached input tokens plus 20 million output tokens. It excludes cache reads/writes, long-context pricing, tool charges, retries, and regional uplifts.
Luna saves $176 and Terra saves $110 for that exact traffic shape, matching the announced 80% and 20% reductions. The arithmetic does not establish equal quality. If a cheaper route adds retries, longer outputs, more tool calls, or human correction, completed-task cost can fall by less—or rise.
Latency premium
Fast mode is not another price cut
OpenAI says GPT-5.6 Sol Fast mode delivers up to 2.5 times faster speeds at twice the Standard price, with no change in intelligence. “Up to” is a provider claim, and actual latency depends on the request and service conditions. Treat the premium as an operational choice: calculate the added API cost, then compare it with the value of time saved.
Fast mode replaces Priority Processing. Existing API requests using service_tier: "priority" continue to work and are routed through Fast mode; OpenAI’s current pricing page says either "priority" or "fast" can be used. That backward compatibility reduces migration urgency, but teams should still update internal documentation, dashboards, and cost labels so operators understand what the old tier now means.
Source-precedence check
Resolve two official-source conflicts before using copied prices
OpenAI’s original July 9 GPT-5.6 launch page still displayed Terra at $2.50 input/$15 output and Luna at $1/$6 when checked August 3. Those figures are launch-era pricing. The dated July 30 announcement, current API pricing page, API changelog, and current Luna model page show the newer rates.
AWS’s July 30 announcement says the new Bedrock rates match OpenAI first-party pricing. The live AWS Bedrock pricing table checked August 3 instead lists Luna at $0.22 input / $1.32 output and Terra at $2.20 input / $13.20 output per million tokens, a nominal 10% uplift over OpenAI Standard. Use the live pricing table for planning, then verify region, private discounts, taxes, and the actual invoice.
- Current provider pricing table for the exact service tier, context band, and route.
- Dated announcement and changelog for the effective change.
- Model page for model-specific context and current quick-reference rates.
- Original launch post only for historical launch-era pricing.
Record the source URL and access date beside any internal cost model, and recheck before a contract, routing, or margin decision.
Five-step routing plan
Reprice first, then test completed-task economics
- 1
Freeze a representative workload
Choose real request shapes with sensitive data removed. Record prompt size, output size, tools, reasoning effort, success criteria, and the current model/service tier.
- 2
Apply the correct rate band
Separate short and long context, Standard/Batch/Flex/Fast, cached reads, cache writes, regional processing, and cloud-provider billing. Do not multiply every token by the headline Standard rate.
- 3
Run Luna and Terra against the baseline
Measure quality, failures, retries, tool behavior, latency, input/output tokens, and human review. A cheaper token is valuable only when the workflow still completes acceptably.
- 4
Price Fast mode against the deadline
Quantify the added Sol API spend and the operational value of lower latency. Keep Standard when the queue already meets the real service-level target.
- 5
Route gradually and preserve rollback
Start with bounded traffic, monitor completed-task cost and errors, and keep the prior route available until the new path clears written quality, latency, and budget thresholds.
What is confirmed and what still requires your test
Effective date, current list rates, long-context bands, service-tier rename, backward-compatible tier value, and the published Fast-mode premium come from current OpenAI sources.
OpenAI’s “up to 2.5x faster” statement and customer performance examples are provider-supplied. TPTK has not independently benchmarked these models or Fast mode.
Quality, actual latency, token use, cache reuse, retries, tool fees, cloud availability, and completed-task economics must be measured in the reader’s environment.
Implementation path
Need a routing and cost evaluation, not another model announcement?
TPTK can help a small business or professional team freeze representative work, compare models and service tiers, measure completed-task cost, and document rollout and rollback rules.
GPT-5.6 pricing FAQ
How much does GPT-5.6 Luna cost now?
OpenAI lists short-context Standard GPT-5.6 Luna at $0.20 per million input tokens, $0.02 per million cached input tokens, $0.25 per million cache-write tokens, and $1.20 per million output tokens. Long-context and other service-tier rates differ.
How much does GPT-5.6 Terra cost now?
OpenAI lists short-context Standard GPT-5.6 Terra at $2 per million input tokens, $0.20 per million cached input tokens, $2.50 per million cache-write tokens, and $12 per million output tokens. Long-context and other service-tier rates differ.
Did GPT-5.6 Sol get a price cut?
No. OpenAI says Sol Standard pricing is unchanged at $5 per million short-context input tokens and $30 per million output tokens. Sol Fast mode is a separate premium tier at twice the Standard token rates.
What happened to OpenAI Priority Processing?
OpenAI renamed Priority Processing to Fast mode on July 30, 2026. Existing requests tagged with the priority service tier remain compatible and use Fast mode; OpenAI’s current documentation permits either priority or fast as the service-tier value.
Is AWS Bedrock the same price as OpenAI direct?
AWS’s July 30 announcement says the new rates match OpenAI first-party pricing, but the live AWS Bedrock pricing table checked August 3 lists nominal Luna and Terra rates 10% above OpenAI Standard. Confirm the live table, region, and your actual account pricing before deciding.
Update log
August 3, 2026 at 12:06 PM PDT: Simplified the first screen, removed a duplicate byline, clarified API savings versus unchanged ChatGPT subscriptions, centered the desktop featured image at a more useful reading width, added the live AWS pricing discrepancy, and replaced the secondary citation with official Fast mode and AWS sources.
August 3, 2026 at 11:05 AM PDT: Initial publication. TPTK rechecked the July 30 announcement, current pricing/model documentation, owner boundary, cost calculations, Image 2.0 assets, category, links, and schema.
Sources
- OpenAI: Advancing the price-performance frontier with GPT-5.6. July 30, 2026; accessed August 3, 2026.
- OpenAI API pricing. Current Standard, Batch, Flex, Fast, context-band, cache, regional, and Bedrock billing notes; accessed August 3, 2026.
- OpenAI Fast mode guide. Service-tier behavior, workload fit, fallback, and dashboard notes; accessed August 3, 2026.
- OpenAI API changelog. July 30 price and Fast-mode update; accessed August 3, 2026.
- OpenAI GPT-5.6 Luna model page. Current model-specific quick-reference rates; accessed August 3, 2026.
- OpenAI GPT-5.6 launch page. July 9 launch-era pricing; accessed August 3, 2026.
- AWS Bedrock pricing. Current nominal Luna and Terra token rates; accessed August 3, 2026.
- AWS: GPT-5.6 Terra and Luna price reduction. July 30 availability and pricing announcement; accessed August 3, 2026.
Every price, availability, and compatibility claim uses an official OpenAI or AWS source. TPTK’s examples multiply disclosed list rates by a disclosed hypothetical workload. No firsthand model-quality or latency benchmark claim is made.
Editorial and monetization note: No approved OpenAI affiliate route or audience-matched Amazon subscription bounty fits this API-pricing decision, so none appears. The implementation link is a TPTK service path. Material corrections and source changes will be recorded in the update log.