▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=vu8X3YroB-w
• Original title :: OpenAI fights back

Theo got early access to GPT-6.1 Soul, shipped one week after GPT-6 Soul, and reads the timing as OpenAI's answer to Opus 5.5. The model is much less erratic than Astra, scores at the top of Terminal Bench and Deep SWE, and is priced so aggressively that it lands near Opus-level results for a fraction of the cost. His verdict: it becomes the best reviewer and investigator in his stack, but Opus 5.5 stays his daily driver for actually writing code.

Pricing is the story

$2 per million input, $10 per million output — same as GPT-6 Soul, half of what 5.6 Soul launched at, a fifth of Astra. The real change is the cache read price: 10 cents per million, a 95% discount instead of OpenAI's historical flat 10%. Since agent workloads are roughly 96% cache reads, this halves real-world agentic cost. Theo argues this cuts into margins (rumoured as high as 95%) and reads Tibo's simultaneous announcement — the $200 Pro sub returning, with usage recalculated to net out at about half the API-dollar equivalent — as confirmation that the subsidisation era is ending.

Concrete cost deltas from his runs: a codebase-improvement task cost $2.15 against $5 for Opus 5.5 and ~$9 for Sonnet 5.5; a Deep SWE run on low matched Opus 5.5 on max for 1/70th the price and 1/10th the time; the priciest 6.1 Soul task was $1.38 versus $5.12 for the cheapest Opus run. Token efficiency compounds it — average context read per request was ~110k tokens versus ~360k for Sonnet.

The spikiness is mostly fixed

Theo's long-standing complaint about Astra was that it was smart and dumb simultaneously, spiking into failures worse than small open-weight models. 6.1 Soul mostly fixes that: lower peaks than Astra, but far more consistent. He trusted it to go through his email, find unpaid invoices, and stage wire transfers in browser tabs — it got nothing wrong. Other early testers agreed; one called it "incredible", another "incredibly boring", which Theo considers the right place to be.

Where it wins and where it doesn't

It excels at deep code review and investigation — the "Rottweiler" behaviour of tearing into a problem until everything wrong surfaces. On a real T3 Code PR it caught a regression that Fable and Opus both missed. It matched Astra's orchestrator audit score at half the price, and placed first in a blind ten-model review benchmark.

It fails at long unattended building. His TypeScript-to-Rust port burned months and hundreds of thousands of dollars of tokens under Astra and Soul with no progress; Opus 5.5 restarted from scratch and got it working in a night for about $1,000. 6.1 Soul then audited that work and found 1.3M of 1.8M lines unused — Opus had rewritten everything in a new crate and left the old slop behind. The models complement each other.

Front-end remains unacceptable. The fish-tank demo looks gorgeous — best fish models and submarine so far — but the movement is janky, the frame rate is worse, and the UI is buried under twenty-plus lines of whimsical filler copy ("Take a breather. Your little world can wait."). Theo says it has regressed, has no taste, and should not be used for design work. It is, however, genuinely good at driving Blender from a screenshot.

Side note: Jev Router

OpenRouter's new cost-optimising Jev Router routed ~60% of requests to DeepSeek v4.1 Flash, ended up more expensive than GPT-6 Astra on low, took 20 minutes against Astra's 4.6, and averaged 104 steps per task versus 19. A model that can't reason and sees no context can't judge problem difficulty.

Takeaway

Theo's plan is to wire 6.1 Soul into Claude Code so Opus 5.5 can call it for investigation, root-cause analysis, and review. He sees no reason to use Sonnet anymore, is done with Astra entirely, and is keeping his Claude subscriptions.