GPT 6.1 Soul: cheaper and smarter, but not a replacement for Opus

The model drop. OpenAI released GPT 6.1 Soul one week after 6.0 Soul, surprising testers with aggressive pricing and competitive benchmark scores. The creator ran Terminal Bench and Deep SWE, finding 6.1 Soul matching Opus 5.5's score on low and high settings at a fraction of the cost—21 cents per task versus $14.65 for Opus 5.5 on max. Cash-read pricing was cut to 5 cents per million tokens, a 95% discount that makes agent workloads dramatically cheaper.

The pricing strategy. At $2 per million input tokens and $10 per million output, 6.1 Soul costs half of 6.0 Soul after discounts and one-fifth of Astra. The creator calculated that his actual usage would have cost $1,550 with this model versus $5,700 at Opus 5.5 prices. This aggressive pricing suggests OpenAI is cutting margins, likely because the model is more expensive to run. The $200 Pro subscription's effective subsidization is dropping, signaling a shift in how OpenAI monetizes access.

Real-world capabilities. On computer use, 6.1 Soul handled invoice-finding and payment setup without errors—tasks the creator wouldn't trust Astra with due to spikiness. Deep code reviews matched Astra's thoroughness at half the price. In a blind test, Opus 5.5 rated 6.1 Soul as frontier-tier for scoped work. The model excels at investigation, bug root-causing, and auditing, making it a strong reviewer.

The spikiness problem. While 6.1 Soul is significantly less spiky than Astra, its occasional dumb mistakes persist. On a long-running Rust-to-TypeScript port, 6.1 Soul burned tokens for four days making no progress. Opus 5.5 restarted from scratch and finished in under a day. For unattended, complex rewrites, Anthropic's model remains superior. The creator still reaches for Opus when collaboration and reliability matter more than cost.

The fish slop test. When asked to build a fish-tank game, 6.1 Soul produced stunning graphics with terrible UX. The UI was cluttered with unnecessary text, performance lagged, and the gameplay loop was broken. This confirms a known weakness: OpenAI models struggle with frontend design and taste. The creator keeps using Opus for code writing and 6.1 Soul for review, planning to set up a workflow where Opus calls 6.1 Soul for investigations.

The verdict for teams. 6.1 Soul is a strong default for scoped, budget-conscious work and excels at code review and investigation. It replaces Sonnet for most tasks. But for complex, long-running projects requiring judgment and reliability, Opus 5.5 remains the better choice. The creator recommends using them in tandem: Opus for implementation, 6.1 Soul for audits. This model changes cost calculations but doesn't dethrone Anthropic for the hardest work.