OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. API pricing is $2 input and $10 output per million tokens for Sol, and $0.10 and $0.50 for Luna. The core headline is a structural price collapse: OpenAI cut API pricing by 50% compared to previous GPT-5.6 promotional rates, with GPT-6 Sol sitting at $2 per million input tokens and $10 per million output tokens, down from $4 and $20. Ninety minutes earlier, Anthropic released Claude Opus 5.5, the first model in its new 5.5 family. The pricing structure is $4/$20 per MTok—a 20% reduction from its predecessor. TechCrunch noted Anthropic released a new version of Opus 5.5 just 90 minutes before OpenAI's release, reflecting the intense competition between the two companies.
I do not run a price war. I run a construction ERP at coenconstruction.com, an estimating SaaS at estimate.pro, and a permit-drawing platform that shipped eight days ago. The ERP calls OpenAI for invoice OCR. The estimator calls Anthropic Claude for line-item extraction from blueprints. The permit tool calls both: Gemini for site-plan markup and Claude for zoning-code summaries. When OpenAI cuts Sol from four dollars to two dollars per million input tokens, my bill goes down. When Anthropic cuts Opus from five dollars to four, the bill goes down again. The invoice parser still fails when the vendor deprecates gpt-5.6-sol and the migration guide says the new model changed how it handles multi-page PDFs. The blueprint extractor still returns hallucinated quantities when Claude decides a dashed line is a wall. The bill dropped. The integration did not.
The models cost less but run the same API
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, just nineteen days after launching flagship GPT-6 Astra, targeting the economic reality of agentic workflows: sustained execution across thousands of tool calls, multi-turn reasoning loops, and production software engineering. OpenAI says the new models were developed using methods similar to Astra's for improvements in professional work, factuality, coding, computer use, and alignment. GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, shows improvements in coding and computer use, and matches or beats Claude Fable 5.1 scores on some benchmarks. The permit platform makes four thousand Claude API calls per month at $0.03 per call. If I switch from Opus 5 at five dollars per million input tokens to Opus 5.5 at four dollars, the monthly bill drops from $120 to $96. If every call finishes in 30% fewer tokens because output speeds increased by over 30%, the bill drops again. If Anthropic deprecates claude-opus-5 on November 1 and the new model requires a different system prompt to achieve the same refusal rate on junk input, I rewrite the prompt, re-run the evaluation set, and redeploy. The token cost went down. The maintenance cost did not.
The construction ERP has three integrations: OpenAI for invoices, Twilio for SMS reminders, and Plaid for bank reconciliation. The estimator has four: Claude for blueprints, Stripe for subscriptions, an SMTP relay for transactional email, and a Cloudflare Worker for PDF rendering. The permit tool has five: Gemini and Claude for document processing, Mapbox for site-plan layers, a Cloudflare D1 database, and a separate Worker for the review gate. Every integration depends on an API I do not control. Twilio can raise SMS prices. Plaid can add a required webhook signature. Stripe can deprecate a checkout parameter. Cloudflare can rename the D1 query method. When the vendor changes the API, I change the integration or the platform breaks. When the vendor cuts prices, the bill goes down and the dependency stays exactly where it was.
The price cut does not reduce vendor risk
Anthropic released Claude Opus 5.5 on September 22, 2026, translating to a 40% reduction in typical workload costs compared to Opus 5 and output speeds increased by over 30%. MacRumors reported OpenAI compared GPT-6 Sol to Opus 5 on several tests, but Anthropic released a new Opus 5.5 model the same day that outperforms Astra on some coding and knowledge work benchmarks. The vendors are competing on price and benchmark scores. They are not competing on API stability, deprecation notice periods, or whether the new model will parse the same invoice the old model parsed yesterday. The price is temporary. The endpoint identifier, the request schema, the system-prompt behavior, and the refusal logic can all change when the vendor ships the next model.
The invoice parser has a fallback. If the OpenAI call fails or returns a refusal, the system logs the failure, emails me the invoice, and lets the user enter line items manually. The fallback exists because I cannot prevent OpenAI from returning an error, changing the API, or deciding that a scanned Sysco receipt looks like a terms-of-service violation. The blueprint extractor has the same fallback: if Claude refuses or hallucinates, the system flags the job and the user reviews it. The permit tool does not have a fallback for the zoning-code summary because there is no manual alternative that runs in two seconds. If Claude refuses to summarize a PDF of the local setback rules, the user does not get a summary. The platform still works. The feature does not.
The model costs half as much per token and still breaks the same way when the vendor deprecates the endpoint.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, with Sol costing $2 per million input tokens and $10 per million output tokens and Luna costing $0.10 and $0.50, both prices half of what OpenAI charges for GPT-5.6 Sol and GPT-5.6 Luna under their current promotional pricing. The token price dropped. The risk that OpenAI deprecates the model, changes the API, or decides my traffic violates a policy I have never read did not drop. The risk that Anthropic rate-limits my API key during a benchmark run against a competing model did not drop. The risk that Gemini returns a 500 error because Google is load-shedding to keep the search product online did not drop. The bill went down. The fleet did not.
I will take the price cut. The estimator will call Opus 5.5 instead of Opus 5 as soon as I verify the new model does not hallucinate rebar quantities at a higher rate than the old one. The invoice parser will call GPT-6 Sol instead of GPT-5.6 Sol after I confirm it still extracts line items from multi-page PDFs the way the ERP expects. The permit tool will stay on Gemini 3.8 Flash until I have time to test whether the API still accepts the same base64-encoded image payload. The price war makes the integrations cheaper to run. It does not make them cheaper to maintain, easier to migrate, or less likely to break when the vendor ships the next model and tells me I have sixty days to switch.