New numbers that stopped me cold: one new AI model hits a 74% coding score for about $2.36 a task, while the model everyone’s hyping does roughly the same work for over $11 a task. Same job. Five times the bill.
I ran into this breakdown from Matt Wolfe, an AI creator who dug into this week’s flood of model releases instead of just cheering them on. He tested three of them himself and pulled the actual benchmark data, and his take flips the hype on its head.
📊 The stat behind the story
Here’s what the original poster found on the DeepSWE coding benchmark:
- Gemini 3.8 Flash: 74% score, about $2.36 per task, 2.5 minutes average
- Claude Opus 5: 74% score, but around $11.84 per task
- Claude Fable 5 (the “best model ever” just six weeks ago): 70%, about $21.63 per task
So a cheap, fast model now codes on par with the state of the art for roughly a tenth of the cost. The creator’s point: because it’s a Google model and Google’s been quiet, almost nobody’s paying attention.
🔍 The insight
The expert argues Fable 5.1 is overhyped and Gemini 3.8 Flash is underhyped. Fable 5.1 does top the intelligence charts (a 66 on Artificial Analysis versus 63 for the previous best), but it’s also the priciest per task at $3.69. To prove the point, he had it build a game clone in one prompt. Gorgeous result. The catch? He accidentally left it on a heavy setting and it burned through $114 in usage. One SVG image alone cost him $4.35 and took 18 minutes.
Meanwhile Gemini 3.8 Flash built its own game clone in a single prompt, ran a test image for 9 cents in 92 seconds, and still looked great.
✅ Three practical takeaways
- Match the model to the job. For everyday coding, the industry pro’s data says a cheap, fast model gets you 90% of the way for a fraction of the price.
- Watch cost per task, not just the headline benchmark. “25% cheaper” claims don’t always survive real workloads, as the creator showed with Fable 5.1 actually costing more per task.
- Don’t sleep on the quiet releases. The person who shared this thinks Google’s model is the real story of the week precisely because it’s not trending.
⚠️ Tips and pitfalls
- Pitfall: leaving a premium model on its heaviest mode. That’s how the author hit a $114 bill on one project.
- Tip: for smart-but-affordable coding, Gemini 3.8 Flash averaged $0.58 per task in his testing.
- Heads up: OpenAI teased an upcoming model called Astra, and the mind behind this video flagged a technique called “recurrent depth” that hides the model’s chain of thought. Great for efficiency, worrying for anyone who needs to audit why a model made a decision, especially in code or security work.
- Bigger picture: the creator admits he’s a bit worn out by weekly “game-changing” drops that only move a benchmark by a fraction. Good reminder to test claims before you trust them.
The honest summary from the one who posted it: the flashiest model isn’t always the smart buy, and the best value this week is the one flying under the radar.
Want the full side-by-side with all his test results and the Astra concerns? Watch the complete video, it’s worth the walkthrough.