Three of the biggest free AI plans changed in three days. OpenAI began moving ChatGPT to GPT-6 on 7 October, Anthropic shipped its cheapest model yet the same day, and Google cut its free Gemini plan to a single model on 9 October. None of it required a decision from the people using them.
The three changes pull in different directions. OpenAI widened access to a newer model. Anthropic made its cheap tier much cheaper and put it in front of free users. Google narrowed its own, limiting free Gemini accounts to Flash-Lite from 9 October.
If you opened ChatGPT on 8 October, or Gemini on 9 October, the model answering you was not the one that answered the week before, and a prompt you saved in September was tuned against something that is no longer there.
OpenAI’s ChatGPT release notes for 7 October are specific: “Plus, Pro, Business, and Enterprise use GPT-6 Sol; Free and Go use GPT-6 Luna.” The rollout started globally that day for the paid tiers in the Chat tab and reached Free and Go the next. Enterprise access depends on workspace admin settings, so an admin can still be holding it back.
Two exceptions matter. The Pro reasoning option “continues to use GPT-6 Astra” and does not support the new interface, and the models behind Work and Codex did not change at all, so somebody paying for ChatGPT Work was running a September model while the free account beside them had the October one.
The release also brings what OpenAI calls Intelligent UI, available from Instant through Extra High reasoning. It mixes text, visuals and interactive elements such as comparisons, diagrams, calculators and games, and picks the format itself. GPT-6 can also start answering while it is still thinking, then add what it finds without a second prompt. Both are presentation changes rather than knowledge: if you have built content workflows that parse a plain text reply, an answer that arrives as a calculator is a new problem.
OpenAI published a system card for both October models on 7 October, and the gap between the free and the paid model has not been this easy to read before. On HealthBench Professional, Sol scores 55.2 and Luna 48.2, both length adjusted. On indirect prompt injection robustness, Sol is at 97.13 percent and Luna at 95.80 percent. Neither of those measures whether a model writes a better subject line, and no published benchmark does. What the card establishes is that OpenAI itself treats the free model as the weaker one on reasoning, safety and tool use.
The more useful finding is in the regressions. OpenAI reports statistically significant regressions for both models on some safety evaluations covering self-harm, gore and sexual content, and on teen safety measures including age-restricted content and emotional reliance. It says the violations it reviewed were borderline and lower severity, and that these evaluations are deliberately difficult and do not reflect typical production rates. Factuality improved on nearly all metrics against the August models, with one exception: claim-level errors on factuality-heavy prompts. If you use ChatGPT to assemble anything fact-dense, that is the line to act on.
Anthropic released Claude Haiku 5.5 on 7 October, describing it as “designed for high-volume, cost-sensitive tasks” such as summaries, classification and database queries. API pricing is $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, against $1.00 and $5.00 for Haiku 4.5. Above 100,000 tokens the rate is $0.50 and $2.50, so crossing that line costs five times more per token.
The threshold is measured on prompt size, not monthly usage, and Anthropic says the cheaper tier covers about 90 percent of the requests that went to Haiku 4.5. It only bites when you paste a long document or push a full transcript through in one go. Anthropic puts the average saving at around 75 percent, with a footnote: 90 percent cheaper below the line, 50 percent above, and the model uses slightly more tokens per task because of an updated tokenizer. The bill falls by less than the per-token cut suggests.
Haiku 5.5 is the first model in its class with an adjustable effort setting, running from Low to Max, so the cost and quality trade sits per request rather than per model. Anthropic’s own table puts it at 1620 on GDPval-AA v2.1 against 1437 for GPT-6 Luna and 1840 for its own Sonnet 5.5. That is the vendor’s chosen comparison, and Anthropic says alongside it that Sonnet 5.5 and Opus 5.5 remain the better models for complex agentic work. Every plan from free up to Enterprise can pick Haiku 5.5 in the web and mobile apps, while Opus 5.5 stays behind a subscription.
Not one of the three vendors published how much you get. There is no stated message count, request cap or rate limit for a free account on any of them, so you cannot work out in advance when a free plan runs out mid-task.
There is also no way to keep the model you had. On a consumer plan you cannot pin last month’s version and there is no opt-out, because pinning a version is an API feature. That is most of the argument for keeping your best prompts in a file outside the app.
Yes, but not the same GPT-6. OpenAI’s release note says Free and Go accounts use GPT-6 Luna while Plus, Pro, Business and Enterprise use GPT-6 Sol. Both are GPT-6 and they are not the same model, and OpenAI’s system card reports Luna scoring lower than Sol on everything it measures for both.
The published numbers cannot answer that for marketing work. The benchmarks the vendors released measure reasoning, coding, safety and professional task scores, none of which tell you which tool writes a better email. The test that does answer it is running your own prompts through each one and reading the output side by side.
No. Haiku 5.5 is selectable on the free plan in the Claude web and mobile apps, alongside the paid tiers up to Enterprise. Opus 5.5, the larger model, is still restricted to paid plans, and Anthropic lists monthly API credit only for Max and Team subscribers.
If the output changed around 8 or 9 October, the model underneath it changed. Rewrite the prompt against the model you are now on rather than reaching for a new tool, and store the working version outside the chat window so the next swap costs ten minutes instead of a week.
Free access to a capable model is wider this week at two of the three vendors and narrower at the third, and in every case the thing you are talking to changed without your involvement. Treat the model as a dependency you check, not a setting you chose. Re-running four prompts is an afternoon for a small team and a quarter for a large company, which is one of the few places you can compete with bigger brands on speed alone.
This is the kind of thing we work on with clients at Social Lady. If you want a second opinion on which AI tools your team should actually be paying for, get in touch.
