Claude Mythos 5 is beating OpenAI’s best model on real software engineering work by 15 points, and that gap is the most honest thing in this release.
Picking the wrong model tier costs teams real money every sprint
Engineering teams running agents at scale have been forced to choose between one model that is too expensive for routine tasks or one that is too weak for complex ones. GPT-5.6 replaces that binary with three tiers priced and tuned for different workloads on the same generation.
Three models, one API, and a runtime the model controls
OpenAI Releases GPT is accessed through the Responses API, where developers select Sol, Terra, or Luna by name and set effort level per call. Sol takes the hardest agentic tasks and outputs structured results at $5 input and $30 output per million tokens; Terra runs at $2.50 and $15; Luna sits at $1 and $6. The sharpest addition is Programmatic Tool Calling, where the model writes and executes its own JavaScript inside an isolated V8 sandbox with no network access, removing the human step of writing tool wrappers entirely.
Agent platform teams feel this before anyone else
- Backend engineers building multi-step pipelines who need parallel agent execution without rewriting orchestration logic
- AI product leads benchmarking model cost against output quality who need a single generation they can tier without switching providers
- Developer tools companies running coding agents who want benchmark-backed bragging rights, given Sol’s score of 80 on the Artificial Analysis Coding Agent Index
The ultra compute mode runs four Sol agents in parallel and lifted Terminal-Bench 2.1 from 88.8 percent to 91.9 percent, which matters most for teams whose workflows have always been the bottleneck, not the model.
OpenAI needed a coding benchmark win and mostly got one
Sol leads the Artificial Analysis Coding Agent Index by 2.8 points over the nearest competitor, but SWE-Bench Pro tells a different story: Sol’s 64.6 percent trails Claude Mythos 5 by roughly 15 points on production-grade software engineering tasks. If that gap closes in GPT-5.7, the competitive calculus for enterprise coding platforms shifts significantly.
What you can actually build with this today
- Run Luna for high-volume classification tasks and Sol only for final reasoning steps
- Set explicit cache breakpoints to control the 30-minute minimum cache life and reduce redundant spend
- Deploy Programmatic Tool Calling to let agents generate and run their own JavaScript logic inside a safe sandbox
- Enable ultra mode to parallelize four agents and push throughput on long-running terminal workflows
Sol is priced at $5 input and $30 output per million tokens; Terra at $2.50 and $15; Luna at $1 and $6, with cache writes billed at 1.25x the uncached input rate and a 90 percent discount on cache reads.
The honest tradeoff: GPT-5.6 Sol trails Claude Mythos 5 by 15 points on SWE-Bench Pro, which means teams whose primary use case is autonomous software engineering should run their own evals before committing.
Claude Mythos 5 still leads on real-world software engineering benchmarks and is worth testing directly if coding agent accuracy is the top priority. Google’s Gemini 2.5 Pro remains competitive on cost-per-output for teams not yet committed to the OpenAI API ecosystem.
The model tier war just moved from marketing to pricing math
Providers are no longer competing on single flagship scores alone — they are competing on whether developers can right-size compute per task inside one platform without context-switching. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.