GPT-6 gets cheaper, Claude Opus 5.5 lands: Sept. 23

OpenAI and Anthropic just made capable AI models cheaper, which gives you a reason to rerun your cost tests before renewing another automation plan. Set aside 30 minutes today to test two model swaps, inspect repeated-input charges, and patch WordPress. You will leave with lower-risk settings, not another tool to buy.

OpenAI cuts GPT-6 Sol and Luna API prices

OpenAI released GPT-6 Sol and Luna on September 22 with API prices 50% below their GPT-5.6 counterparts. The API is the connection that lets another piece of software send work to the model. Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 and $0.50 respectively. Tokens are the small text units used for billing.

That price gap changes model routing. Use Luna for repetitive sorting, tagging, and first drafts. Reserve Sol for jobs where the model must reason through messy instructions or use several tools. Do not replace every model at once. A cheaper wrong answer still wastes staff time.

Open your automation tool, duplicate one routine, and change only its model name to gpt-6-luna. Run 20 saved examples through the copy. Compare its answers, response time, and estimated token cost with the live version. For more guarded testing, use the review pattern in the AI content approval workflow.

Claude Opus 5.5 lowers the cost of difficult work

Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens. Anthropic says those rates are 20% below Opus 5, while typical workloads cost 40% less because reused text is cheaper. The new model is available through Anthropic and major cloud providers under claude-opus-5-5.

This is not the model for bulk email classifications. Its value is difficult, supervised work where a weaker model creates expensive rework: checking a long contract, tracing a software fault, or reconciling conflicting research. Keep a person responsible for the final decision.

Duplicate one demanding Claude routine and replace only the model identifier. Use five completed jobs with known-good answers, then record missed requirements and total cost beside Opus 5. If you rely on Claude during outages, add the fallback and alert pattern from yesterday's dispatch.

OpenAI exposes prompt-cache misses

OpenAI also added a Prompt Caching Dashboard and a cache-miss diagnostic on September 22. Prompt caching means the service can reuse repeated instructions instead of processing the same text again. Eligible shared text reused within 30 minutes can receive discounts of up to 90% on input-token charges.

This matters when an automated assistant sends the same policy, product catalog, or tool list with every request. Changing the order of those instructions can force the service to process them again, raising cost and slowing replies.

In the OpenAI Platform, open Usage, then Prompt Caching. Find your busiest GPT-6 project and note its cached versus uncached input. If cache use drops after a workflow edit, run the diagnostic and look for tools_changed, a model change, or reordered instructions. Keep the stable material first and put the customer's new message last. The same cost check belongs in any weekly AI marketing report.

WordPress patches a critical page-template flaw

WordPress published a critical security advisory on September 22 with a 9.2 severity score and patched every supported branch back to version 4.7. Under specific theme and server conditions, an attacker without an account can make WordPress load a PHP file outside the theme folder. PHP is the software language that runs much of WordPress; loading the wrong file can let an attacker take control of the site.

If your website collects inquiries, this outranks model testing. A compromised form can expose customer details and break every follow-up attached to it.

Open WordPress Dashboard → Updates and install the offered core security release. Version 7.1 users need 7.1.2; older branches have their own patched release. Back up the files and database first, update, then submit a test form and confirm the lead reaches the software that stores your customers and sends follow-ups. If that handoff uses a web address that catches submitted data, rerun the checks in the Make webhook setup.

On the bench

• Test GPT-6 Sol only on work where Luna misses a written requirement; price alone is not a reason to use the larger model.
• Watch Opus 5.5's higher plan limits before changing any team allowance. Usage capacity is useful only when the review process can keep up.
• Add prompt-cache use to your next monthly cost review, alongside model price and failed jobs.

Get the next build in the AI tools runbook.