OpenAI has launched GPT-6 Astra, its latest frontier mannequin, lower than every week after Anthropic’s Claude Fable 5.1. OpenAI calls Astra the world’s most clever and aligned mannequin but.
What truly makes Astra completely different?
The best technique to put it’s this: Astra is constructed to do extra, not simply reply extra. It might use a pc to finish duties as a substitute of telling you the best way to do them. It might create completed paperwork as a substitute of providing you with a tough first draft. Throughout lengthy coding classes, Astra may bear in mind what occurred earlier as a substitute of ranging from scratch. And it’s higher at figuring out when to take motion and when to cease and ask.
The larger shift isn’t simply smarter solutions. AI is getting nearer to really getting the work accomplished.
Some essential caveats stay, although. Astra doesn’t beat each competing mannequin, and a few of its greatest benchmark numbers include asterisks. Let’s have a look at what’s new, what Astra does effectively, and the place the caveats are.
What Truly Modified in GPT-6 Astra
Drive a Laptop Finish-to-end
Astra can fill out types, replace CRM data, run frontend QA checks on an internet site, and troubleshoot software program by watching what occurs on display, with out step-by-step hand-holding. OpenAI studies a rating of 72.6% on OSWorld 2.0, a benchmark for actual desktop pc use, narrowly forward of Claude Opus 5’s 70.2% and effectively forward of GPT-5.6 Sol’s 65.7%.
Instance: GPT-6 Astra can extract info from a W-2 and enter the related particulars right into a Kind 1040, serving to automate routine tax-preparation work whereas leaving the ultimate return for human evaluate.
Is aware of When to Ask and When to Guess
Earlier fashions tended to both guess mistaken on ambiguous directions or interrupt with pointless questions. Astra is educated to fill in routine gaps by itself and pause solely when the reply would meaningfully change the end result.
OpenAI’s personal side-by-side demo reveals the distinction: GPT-5.6 Sol constructed a private profession web site by itself in about 13 minutes. Astra paused after 20 seconds to ask what profession the consumer was truly transferring into. That’s a small second, nevertheless it’s a very good illustration of the distinction between a mannequin that acts and a mannequin that makes use of judgment about when to behave.
Produces Completed Paperwork
There’s a significant distinction between an AI that generates content material and one which completes a deliverable. A generic mannequin palms you textual content for a presentation. OpenAI says Astra is educated to match your present templates, tone, and construction, and to tug in solely the context that’s related moderately than padding the output with all the pieces it is aware of.
Astra builds a slide deck from only a handful of template slides whereas holding the tone and format constant all through. For groups that at the moment spend time reformatting AI output right into a home fashion, that’s the half price testing first.
Remembers Throughout Lengthy Coding Periods
In Codex, Astra can now hold searchable notes throughout context home windows as a substitute of repeatedly compressing lengthy debugging classes right into a single abstract, a change OpenAI says preserves particulars that compaction tends to lose, like why an earlier repair failed. It’s opt-in for now by Codex’s config file, and OpenAI says it can develop into the default within the coming weeks.
Cybersecurity
Astra’s greatest soar isn’t a productiveness function in any respect. OpenAI studies Astra reaches the “Important” threshold for cybersecurity below its personal Preparedness Framework, its highest threat tier, which means the mannequin can independently determine and develop working exploits for beforehand unknown vulnerabilities. OpenAI studies a 100% rating on ExploitBench and says Astra solved 88% of SRE-Bench reverse-engineering duties on the primary try.
Due to that threat tier, OpenAI is gating the extra harmful elements of this functionality at launch. Astra will assist with defensive work like safe code evaluate and patch validation, nevertheless it refuses to create proof-of-concept exploits till OpenAI expands entry by its Dawn program.
GPT-6 Astra Benchmarks
The desk beneath pulls collectively the headline comparisons OpenAI printed towards GPT-5.6 Sol, Claude, and Gemini. All figures are self-reported by OpenAI in its launch supplies.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 | Claude Opus 5 | Gemini 3.8 Flash |
|---|---|---|---|---|---|
| OSWorld 2.0 (pc use) | 72.6% | 65.7% | – | 70.2% | – |
| FrontierMath Tier 4 | 97.6% | 83.0% | 87.8% | 73.2% | – |
| GPQA Diamond | 96.0% | 94.6% | 93.7% | 93.7% | 95.3% |
| Terminal-Bench 4.0 (coding) | 57.7% | 37.3% | 55.8% | 52.3% | 19.1% |
| ExploitBench | 100.0% | 78.5% | – | 70.0% | – |
| Humanity’s Final Examination (w/ instruments) | 57.2% | – | 65.0% | 63.6% | – |
What this truly tells us:
- Astra’s clearest energy is pc use. It leads Claude Opus 5 on OSWorld 2.0 and does it in meaningfully much less time per process.
- Its coding scores have improved considerably over its personal predecessor, although the lead over Claude Fable 5.1 on Terminal-Bench 4.0 is slim moderately than decisive.
- It’s not universally forward of Claude. On Humanity’s Final Examination with instruments, Astra scores 57.2%, behind each Claude Fable 5.1 (65.0%) and Claude Opus 5 (63.6%).
- Some headline numbers depend on analysis setups that don’t mirror regular utilization. Astra’s marketed 99.9% on ARC-AGI-3 is dependent upon an costly, stateful analysis harness. Unbiased testing by the ARC Prize Basis discovered that a typical, stateless API name scores far decrease, someplace between 17% and 63% relying on the reasoning tier used. Anybody calling the mannequin by a traditional API integration ought to count on the decrease finish, not the headline determine.
Additionally Learn: GPT-5.6 Sol vs Claude Fable 5: Benchmarks, Pricing & Palms-On
GPT-6 Astra vs Claude Fable 5.1 Comparision
As neither mannequin is publicly out there but, I turned to X to see what folks with early entry are sharing about their experiences. Within the following part, I’ll spotlight among the most attention-grabbing examples and head-to-head movies I got here throughout.
Constructing a 3D Villa from a Single Immediate
Developer Karan Kendre posted a head-to-head Blender comparability, asking Claude Fable 5.1 and GPT-6 Astra to construct the identical villa scene. Astra comes out forward within the visible comparability, with a extra polished and realistic-looking end result, significantly within the inside particulars and general presentation.
Supply: KaranKendre on X
Designing a Journey App from the Similar Immediate
On this process GPT-6 Astra and Claude Fable 5.1 have been prompted to design an app primarily based on the identical immediate and objective. Fable went with a sky-themed design, whereas Astra took an astro-inspired strategy. Astra comes out forward right here, with a extra distinctive visible identification and stronger thematic consistency throughout the design, whereas Fable’s model feels cleaner however extra standard. It’s a very good instance of how two fashions can interpret the identical inventive transient in very alternative ways.
Supply: jaimin on X
Video Immediate Technology
GPT-6 Astra and Claude Fable 5.1 have been every used to generate a immediate for a similar video idea, which was then created utilizing Higgsfield AI. Astra comes out forward right here, with a extra coherent sequence and stronger visible storytelling. Its model maintains the monk, big koi, and motion extra constantly throughout photographs, whereas Fable’s model feels extra dramatic however has some much less constant parts towards the top. It’s a very good instance of how the standard of the immediate can affect the ultimate AI-generated video.
Supply: HiggsField AI on X
Value of GPT Astra
GPT-6 Astra is rolling out in levels: beginning with a restricted set of organizations, then increasing to all ChatGPT Plus, Professional, Enterprise, and Enterprise customers over the next days. Enterprise admins have to manually allow it for his or her workspace, because it’s off by default at launch. Professional, Enterprise, and Enterprise customers additionally get entry to a separate GPT-6 Astra Professional tier.
For builders, Astra is accessible as gpt-6-astra by the OpenAI API, Microsoft Azure, and Amazon Bedrock.
API pricing:
- Enter: $10 per million tokens
- Output: $50 per million tokens
- Quick mode: roughly 2.5x the velocity of ordinary processing, at 2x the value
That’s effectively above GPT-5.6 Terra’s $2/$12 charges and Claude Opus 5’s $5/$25 charges. The pricing itself is a sign: that is constructed for high-value autonomous work, not routine bulk-text technology. Groups paying $10/$50 per million tokens are doubtless doing so as a result of a process will get completed reliably with minimal supervision, not as a result of they want extra phrases generated per greenback.
The mannequin additionally helps zero information retention for eligible API clients, and OpenAI says it’s testing non-public security processing to strengthen monitoring w hile preserving buyer privateness.
Be taught extra about GPT-6 Astra here.
Conclusion
The actual check for Astra received’t be whether or not it may possibly high one other benchmark. It is going to be whether or not corporations can hand it a messy, multi-step process and belief it to get the job accomplished with minimal supervision.
On paper, Astra appears to be like like a big step ahead, significantly in pc use, coding, and agentic workflows. However the benchmarks additionally present that it doesn’t lead in all places. Claude’s fashions nonetheless have an edge on broader reasoning benchmarks, which is why the appropriate mannequin in the end is dependent upon the work you want it to do.
The larger query is what occurs exterior the leaderboard. As soon as Astra is broadly out there, real-world testing will present whether or not its strengths translate into dependable, on a regular basis workflows. That’s the check that issues most, and one solely hands-on use can reply.
Ceaselessly Requested Questions
A. GPT-6 Astra is OpenAI’s latest frontier mannequin, succeeding GPT-5.6 Sol. OpenAI positions it round three strengths: pc use, producing completed skilled work, and a significant soar in cybersecurity functionality.
A. OpenAI studies Astra outperforms Sol on most measured benchmarks, together with a considerable lower in time per process on pc use, a big soar on math benchmarks, and an ideal self-reported rating on ExploitBench. It’s additionally OpenAI’s first mannequin to cross the Important cybersecurity threshold below its Preparedness Framework.
A. Astra studies a lead over Claude Opus 5 on pc use and a slim one on coding, however trails Claude Fable 5.1 and Opus 5 on Humanity’s Final Examination with instruments. It’s a robust mannequin with clear specializations moderately than one which wins throughout each metric.
A. It’s rolling out to ChatGPT Plus, Professional, Enterprise, and Enterprise customers, together with the OpenAI API, Microsoft Azure, and Amazon Bedrock. Enterprise workspaces have to manually allow it, because it’s off by default at launch.
A. Commonplace API pricing is $10 per million enter tokens and $50 per million output tokens, with quick mode out there at roughly 2.5x velocity for 2x the value. That’s notably costlier than Claude Opus 5’s $5/$25 charges.
A. OpenAI classifies Astra as reaching the Important threshold for cybersecurity threat, so its most superior exploit-creation capabilities are gated behind the corporate’s Dawn program at launch. It ships with further alignment coaching and safeguards, although OpenAI has flagged a regression in how simply its reasoning might be monitored, and lists that as an ongoing analysis precedence.
Hi there, I’m Nitika, a tech-savvy Content material Creator and Marketer. Creativity and studying new issues come naturally to me. I’ve experience in creating result-driven content material methods. I’m effectively versed in search engine marketing Administration, Key phrase Operations, Net Content material Writing, Communication, Content material Technique, Enhancing, and Writing.
Login to proceed studying and luxuriate in expert-curated content material.
