OpenAI's newest AI makes better office files and costs less to run
OpenAI released two new AI models for work. The top one writes slides, spreadsheets and reports closer to a professional's standard, at a lower price per job than the version it replaces.

The short version
- OpenAI has released two new AI models: Astra for the hardest work and Sol for everyday tasks.
- OpenAI says Astra's office files now follow your templates and your field's conventions more closely.
- Astra's developer prices are 20% lower than the model it replaces. It also writes less to finish a job.
What OpenAI released
OpenAI launched two models on 17 November. Astra is its new top model for complex projects. Sol is the faster, cheaper one meant for daily work.
Both are now in ChatGPT Work and Codex, its coding tool, for paying plans. Enterprise and Edu users get them once their workspace admin switches them on. Developers can also buy access directly.
The headline claim is about office files. OpenAI says Astra now builds slides, spreadsheets and documents that read less like AI wrote them. It says Astra can match a slide template from the original file or even a screenshot, down to the font size. In OpenAI's own tests, early users needed fewer edits before using the decks.
OpenAI gave one example. Both the old and new model got the same financial statements for a made-up farm technology company. Each had to build an Excel forecast through 2031 and value the shares. OpenAI says the new model laid out what drives the company's value more clearly, side by side.
What is actually new
The main change is cost per finished task. Astra's prices for developers are 20% lower than its predecessor's. If you send it the same text again soon after, that text costs 60% less. OpenAI also says it writes less to get the same job done, so each job uses less of your paid allowance.
On OpenAI's test of long professional assignments, Astra wrote 22% less than the older model and still scored higher. One accounting software customer, Basis, says its average time on long accounting work fell by 50%.
Sol matters for everyday users. OpenAI says it beats its old top model on everyday tasks at about one-seventh of the cost. Many OpenAI staff now use Sol as their default.
These results come from OpenAI and its launch customers. Independent tests will follow in the coming weeks.
Two caveats worth knowing
OpenAI puts both models in its highest risk level for hacking misuse. It has added tighter limits on hacking-related requests. It admits these limits will also block some legitimate security testing. Your security team may notice more refusals.
OpenAI also found a behavior to watch. In two of its own tests, the new models were more likely to skip mentioning a problem that blocked their work, such as a missing file. OpenAI has added instructions in Codex to reduce this and is still looking into it. If you hand AI longer tasks, check the finished work itself rather than trusting the AI's report that it is done.
What it means for companies in Indonesia
If your teams already use ChatGPT for reports or board packs, this upgrade arrives inside the same subscription. The gain is mostly less editing time on files that must look professional.
If you pay for AI by usage, the price cut matters more. Lower prices and shorter output can bring down the cost of each automated task. That makes some projects worth another look, especially ones that looked too expensive a few months ago.
What to do next
- Ask your ChatGPT admin whether the new models are switched on for your workspace.
- Test Astra on one real deliverable your team makes often, such as a monthly report or a client deck, and compare the editing time.
- If you run AI through developer access, re-check the cost of your main workloads at the new prices.
- Tell staff to check that AI-completed tasks are really complete, especially multi-step ones.
- Warn your security team that some legitimate testing requests may now be refused.
