Three AI labs shipped their biggest models within four weeks of each other, and every one of them claims to be the best. OpenAI released GPT-6 Astra on September 3. Anthropic released Claude Opus 5.5 on September 22. Google announced Gemini 4 Argon on September 30 and opened it widely on October 13.
Launch blogs tell you what a model can do. They rarely tell you what happens on a Tuesday afternoon when a developer, a security team or a small startup actually puts it to work.
So this piece skips the press releases and looks at the people using these models right now: what they asked for, what they got back, and whether they were happy with it.
GPT-6 Astra: brilliant at doing, shaky at following through
Astra’s early users are thrilled when it drives software for them, and frustrated when it has to finish the boring last mile. It costs $10 per million input tokens and $50 per million output, the same headline price as Claude Fable 5.1.
The product builder who finally cracked a six-month problem. Claire Vo, founder of ChatPRD, says Astra one-shotted a product-intelligence feature she had failed to build with GPT-5.6 Sol and Fable. She also used it to crack a hardware hack she had chased for months and to build 3D assets in Blender (Lenny’s Newsletter). Verdict: very satisfied.
The investor who made it his daily driver. Matt Shumer says Astra “won me back” and is now his default for most work. It reviewed his tax information and found possible savings his accountants had missed. But a website redesign he asked for came out poorly, and on very long projects progress still slows down (Something Big). Verdict: satisfied, with limits.
The code-review company that measured it. CodeRabbit found Astra caught about 4% more labeled bugs than GPT-5.6 Sol and 22% more than Claude Opus 5. Its team also had Astra build and ship a full co-op game, handling Xcode, certificates and notarization on its own (CodeRabbit). Verdict: satisfied.
The veteran engineer who went back. Armin Ronacher, creator of the Flask framework, called Astra amazing but said he returned to GPT-5.6 for day-to-day software engineering. He described it as the first OpenAI release that felt like a regression for his workflow. Verdict: not satisfied for coding.
The YouTuber who caught it skipping steps. Theo of t3.gg had Astra fix a performance problem, and it opened a pull request within 10 minutes. When review bots flagged issues, Astra agreed they needed fixing, then fixed them without pushing the changes. He still prefers Anthropic’s Fable 5.1 for front-end design. Verdict: mixed.
The world builders. Game and 3D creation became the launch’s most popular use. Matt Shumer gave Astra a week inside Unreal Engine and it built a replica of Manhattan. Designer Tom Krcha handed it one photo of a house and got back the full interior as editable 3D geometry running at 60 frames per second, down to the appliances and toys. Verdict: blown away.
The ambitious side project. Shumer also had Astra build a working web browser. It wrote its own tests, compared its browser against Chromium and kept fixing the differences on its own. This is the long, self-checking work Astra was designed for.
The everyday errand. OpenAI’s own example is the most relatable: an apartment search that would take a person about six hours, done by an Astra agent in under 10 minutes. Real users have not yet published independent timings for this, so treat it as a demo, not a guarantee.
The pattern, as Decrypt put it after a weekend of tests: Astra shines on work with a checkable right answer, like code that compiles or a 3D mesh that renders, and is only average where taste decides (Decrypt).
Claude Opus 5.5: the cheaper model that won people back, if you set a budget
Opus 5.5 is winning over users who had given up on Claude, mainly because it delivers near-flagship work at a lower price. It costs $4 per million input tokens and $20 per million output, against $10 and $50 for Anthropic’s top model, Fable 5.1.
The consulting giant’s code reviews. Deloitte’s CIO says Opus 5.5 at its lowest effort setting caught 72% of known bugs, compared with 56% for Opus 5 at high effort (CodingFleet). Verdict: satisfied.
The 20-hour job done in three. Anthropic cited an early tester who audited and fixed a 200,000-line codebase in under three hours, a job that took more than 20 hours with Opus 5, while using 2.5 times fewer tokens. Verdict: satisfied, though this is a vendor-selected example.
The engineer who switched his daily driver. At Every, Kieran Klaassen moved his daily coding from Fable 5.1 to Opus 5.5. On one Ruby task it wrote 14 clean lines where Fable wrote 26, and he rates it about 90% as capable as Fable for code (Every). Verdict: very satisfied.
The designer who came back. Tyler Nishida, a designer at Every who had cancelled his $200-a-month Claude plan, said his “jaw dropped at least five or six times” in a week. In his tests it followed visual style rules better than Astra and Fable. Verdict: very satisfied.
The evals lead who got burned twice. Every’s head of evals, Mike Taylor, asked for a voice-driven form app. After 30 minutes and about 5.9 million tokens, its automated checks passed, but the core screens threw errors when he actually used it. Given 10 minutes to write a client training schedule, it built extras and ran out of time without delivering the schedule. Verdict: impressed by its thinking, not by its finishing.
The code-review team that counted the trade-offs. CodeRabbit found Opus 5.5 caught 10 of 13 harder bug cases versus 5 for its production setup. The cost: roughly 50% more tokens, and 9 bugs its baseline caught that Opus 5.5 missed. Verdict: mixed.
GitHub’s Copilot team. GitHub’s chief product officer says Opus 5.5 used among the fewest tokens and steps they measured across Copilot CLI and VS Code. In VS Code it solved more terminal tasks than Opus 5 in less than half the steps. Verdict: satisfied.
Box’s document AI. Box runs AI over huge amounts of customer content, so speed and cost matter most. It reports Opus 5.5’s answers were 40% less wordy. Verdict: satisfied.
The finance analyst test. In one test, both Opus 5.5 and Opus 5 built an Excel model of a fictional merger and turned it into an executive deck. They reached the same conclusions, but Opus 5.5’s model was more thorough, its deck easier to read, and it finished in 63 minutes instead of 93, at half the cost.
The big rewrite. Rewriting HAProxy, a widely used load balancer, from C to Rust, Opus 5.5 passed nearly all regression tests in 9.5 hours. Fable 5.1 took 12 hours and cost about twice as much.
The writers. Every’s team measured Opus 5.5’s prose as the most readable they have tested from Anthropic or OpenAI. But in one essay test it took 37 to 39 sentences to say what the human writer said in 21. Verdict: a pleasant collaborator that still buries the point.
One more catch for security teams: Opus 5.5 ships with stricter safeguards, and most cybersecurity tasks are redirected to the older Opus 4.8.
Gemini 4 Argon: a great writer and editor, with a coding debate
Argon has put Google back in the top tier, especially for writing and knowledge work, but its coding is the most argued-about of the three. Until December 1 the API is half price, at $2 per million input tokens and $10 per million output (Techzine).
The enterprise content company. Box CEO Aaron Levie says Argon scored 14 percentage points higher than Gemini 3.1 Pro on Box’s evaluation of complex work. Verdict: satisfied.
The writer who saved two to three days. Every CEO Dan Shipper calls Argon the best structural editor he has used. It cut two to three days off an 8,000-word writing project. The same panel found it weaker than Claude at checking and revisiting its own code, and said Google’s apps hold the model back (BigGo summary). Verdict: loves the model, not the app.
Google’s own engineers. Bloomberg reported that some Google employees found Argon worse at real-world coding and front-end design than its benchmarks suggest. Google disputes this, pointing to a 35% rise in agentic code pull requests and nearly a third fewer rollbacks internally. Verdict: divided.
Everyday users on Hacker News. Some call it exceptional for writing and prefer its directness to Opus 5.5. Others say it still makes things up, and one user said Gemini wiped their chat memory eight times before they mostly stopped using it. Verdict: mixed.
The leaderboards. Argon ranks first on the Vals Index and first in LMArena’s blind text test, including its creative writing category. On the Artificial Analysis index it scores 53, level with Astra but behind Opus 5.5 at 58. Which leaderboard you trust changes who “wins”.
The hands-on testers. Analytics Vidhya ran Argon through four practical tasks. It built a working pottery workshop booking site in Gemini’s Canvas and handled a tricky policy-document task correctly, but it also slipped an unsupported claim into a launch summary. Verdict: useful, but check its facts.
The 3D and science crowd. Early users praised Argon’s 3D generation; one built a realistic 3D library from a single prompt. Side-by-side tests found its games playable and its physics working, but less polished visually than Opus 5.5 and Astra. Verdict: good, not the best, at visuals.
The “I don’t know” trade-off. Argon has the lowest measured hallucination rate among top models, at 15%. But it answers only 50% of hard factual questions correctly versus 63% for Astra, because it declines more often. One reviewer also found it uses about 62,000 output tokens per task against Astra’s 27,000, which pushes its cost per task to about $3.98 once the discount ends. Verdict: safer answers, bigger bills.
The red flags users keep running into
Across all three models, the same four complaints come up again and again. None of them shows up in benchmark tables.
“Done” that isn’t done. Astra fixed review comments but never pushed the fixes. Opus 5.5 passed its own checks on an app whose main screens were broken. Both look finished until a human opens them.
Running past the brief. Opus 5.5 spent nearly two hours on a one-prompt golf game and missed a 10-minute deadline by building extras nobody asked for. Every’s testers hit their weekly limits during a single week of testing.
Confident mistakes. Argon added an unsupported claim to a summary, and some Hacker News users say Gemini still invents answers that a quick search would correct. Argon’s low hallucination rate comes partly from declining to answer more often.
New is not always better. Armin Ronacher went back to GPT-5.6 for coding. Some Google engineers reportedly found Argon weaker on real code than its benchmarks suggest. A model that tops a chart can still feel like a step back in your own workflow.
The fix is the same in every case: give clear limits, check the output yourself, and keep the previous model handy until the new one proves itself on your work.
The scorecard: who got what they wanted
Of the 16 main cases above, 9 came away satisfied, 5 were mixed and 2 were let down on the task they cared about.
| User | Model | What they did | Outcome | Satisfied? |
|---|---|---|---|---|
| Claire Vo, ChatPRD | GPT-6 Astra | Built a feature she had failed at for six months | One-shotted it | Yes |
| Matt Shumer | GPT-6 Astra | Daily work, tax review, website redesign | Found missed tax savings; redesign was poor | Yes, with limits |
| CodeRabbit | GPT-6 Astra | Code review and a full game build | ~4% more bugs caught than GPT-5.6 Sol | Yes |
| Armin Ronacher | GPT-6 Astra | Everyday software engineering | Went back to GPT-5.6 | No |
| Theo, t3.gg | GPT-6 Astra | Performance fix via pull request | PR in 10 minutes, but fixes never pushed | Mixed |
| Deloitte | Opus 5.5 | Code review | 72% of known bugs caught vs 56% before | Yes |
| Early tester (via Anthropic) | Opus 5.5 | Audit of a 200,000-line codebase | Under 3 hours vs 20+ hours | Yes |
| Kieran Klaassen, Every | Opus 5.5 | Daily coding | Replaced Fable 5.1 as his main model | Yes |
| Tyler Nishida, Every | Opus 5.5 | Design and visual prototypes | Beat Astra and Fable on style | Yes |
| Mike Taylor, Every | Opus 5.5 | App build and timed client deliverable | App broke; deadline task missed | No |
| CodeRabbit | Opus 5.5 | Hard code-review cases | 10 of 13 caught, but ~50% more tokens | Mixed |
| Box | Gemini 4 Argon | Complex enterprise work | +14 points vs Gemini 3.1 Pro | Yes |
| Dan Shipper, Every | Gemini 4 Argon | Editing an 8,000-word project | Saved 2 to 3 days | Yes |
| Google engineers | Gemini 4 Argon | Real-world coding | Some say worse than benchmarks | Mixed |
| Hacker News users | Gemini 4 Argon | Writing and daily chat | Strong writing; memory and accuracy gripes | Mixed |
| eesel reviewer | Gemini 4 Argon | Cost per task analysis | ~2.3x the output tokens of Astra | Mixed |
What this means for you
You don’t need to be a developer to benefit, but the right pick depends on what you do all day. Based on the experiences above, here is where each model has helped people most.
| If you are a… | Start with | Why, based on real users | Watch out for |
|---|---|---|---|
| Student or job seeker | Gemini 4 Argon | Strong structural editing for essays and long documents; half-price API until December 1 | Check every fact and citation it gives you |
| Writer or marketer | Gemini 4 Argon or Opus 5.5 | Argon for restructuring long drafts; Opus 5.5 for the most readable first drafts | Both may bury your main point; move it to the top yourself |
| Designer or creator | Opus 5.5 | Beat Astra and Fable at following visual style rules in Every’s tests | Long, costly runs if you don’t set a limit |
| Developer | Opus 5.5, with Astra nearby | Cheaper daily coding and strong code review; Astra for agent work that drives other apps | Run the app and read the changes before shipping |
| Business owner or analyst | GPT-6 Astra | Best at operating software for you, from spreadsheets to browsers | Higher price per task; safety checks can pause work |
A simple way to decide: take one real task from last week, run it through two models, and compare the time, the cost and how much you had to fix. That one test will tell you more than any leaderboard.
So which one should you use?
The honest answer from real users is: the one that fits the job, checked on your own work. Each model has a clear sweet spot.
- Pick GPT-6 Astra when you want an agent to drive software for you, like browsers, design tools or game engines, and the result can be checked.
- Pick Claude Opus 5.5 for daily coding, design and creative work where cost matters, but set a time and token budget and test the output yourself.
- Pick Gemini 4 Argon for writing, editing and long documents, and use the half-price window until December 1 to try it cheaply.
The biggest lesson from these 16 stories is not about any one model. The happiest users checked the work before trusting it. The unhappiest ones trusted a green checkmark that turned out to be wrong.
Want the next breakdown before everyone else?
New AI models now arrive every few weeks, and every launch comes with big promises. We dig into what real people actually get from them, in plain English, with real numbers and no hype.
Follow us so the next breakdown lands in your feed first. Coming up next: we give all three models the same real-world task and compare the time, the cost and the final result side by side.
Now we want to hear from you. Which of these models have you tried, and did it live up to the hype? Share your experience in the comments, and we’ll feature the most useful answers in our next post.
If this saved you hours of research, pass it on to one friend or teammate who is still deciding which AI to pay for.
