▶ YouTube Summaries

• https://www.youtube.com/watch?v=IyoTJHLmClo
• Original title: GPT-5.6: The Review

Theo's full review of OpenAI's GPT-5.6 family — Soul (flagship), Terra (balanced), and Luna (cheapest) — combining benchmark numbers, aggregated reactions from other well-known developers, and guidance on navigating the 30-plus combinations of model, reasoning effort, Pro, and Ultra. His verdict: 5.6 Soul is a phenomenal, state-of-the-art and unusually token-efficient model that has become his default for most work, with real weaknesses around over-eager code generation, stubbornness, and runaway token burn. He defers the Soul-vs-Fable comparison and a deep Ultra breakdown to later videos.

Benchmarks and OpenAI's claims

5.6 is a refinement (post-training + RL) of 5.5, not a new base model — which makes its jump striking. On DeepSWE, Soul on max set a record 73% at ~$22/task versus Fable's ~$839 max-equivalent (Soul scoring higher and far cheaper). On the Agents' Last Exam (long-running professional workflows across 55 fields), Soul set 53.6, beating Fable 5 by 13 points; even at medium it beats Fable by 11 at ~a quarter the cost, and notably Soul on XHigh (~$760) outperformed Soul on max. On the Artificial Analysis coding agent index, Soul on max leads by ~20 points (next: Fable 77, Grok Build 76), with Terra tying Fable. Soul with max comes within one point of Fable on the general intelligence index while finishing 61% faster at ~half the cost. Terra repeatedly lands near or below 5.5's cost while scoring higher, which Theo calls underrated. A key capability: 5.6 can write and run lightweight programs that coordinate tools (programmatic tool calling via the responses API), filtering large intermediate data so tool-heavy tasks use fewer tokens and round-trips. GLM-5.2 illustrates his recurring point that a low per-token price still loses on total cost when the model is token-inefficient.

Design, safety, rollout

Design is better than 5.5 (a low bar) but still not frontier — without careful steering it produces slop; with steering and the design skill it can be decent, and it's surprisingly good at 3D. It likes a "marquee" style and needs handholding, but it listens well when corrected. Cyber/science scores are strong (ExploitBench, GPQA-Pro, life-science benches), though allegedly still behind the unreleasable Mythos. Safety is aggressive: 5.6 blocks ~10x more potentially harmful activity than prior models, creating friction for benign use (Theo had a legit refactor blocked); there's a retry-on-lower-capability-model option. Internally, OpenAI's research compute for coding inference grew ~100x and agentic token usage ~22x over six months; per-researcher daily output tokens more than doubled versus 5.5. Rollout: Soul to paid tiers, Terra/Luna to free-and-go; API pricing ~$5/$30 per million in/out for Soul, ~$2.50/$15 for Terra, ~$1/$6 for Luna, now with billed cache rates.

What other developers say

Many early testers had access since May 27, lost it at announcement, and were "depressed" without it — falling back to older models felt terrible, which clarified how much better 5.6 is. Dax: never hyped a release before, team using 5x the tokens, "not even smarter than Fable, just so reliable and fun." Max: its defining trait is that it never gives up — on max reasoning it keeps working until done. Mitchell Hashimoto: made Soul his default; faster, plans/judges as well as Fable, better overall work. Tim (Next.js): two months of use, understands architecture trade-offs, investigates complex issues, needs little guidance, landed big end-to-end server refactors. The Every team and others echo: fast, finds its own context, one thread can carry a project, compaction actually works, but it may build too much and works best when you plan to steer it. Cory's sarcastic jab (unclear price-hike date, subscription removal, docs/employee misalignment) is a burn on Anthropic; an OpenAI employee confirmed 100% of quota on 5.6 forever at unchanged price.

Strengths

Determined to a fault (finds a way to complete any completable task), better front-end than 5.5, industry-leading computer use, very token-efficient and fast (a Cerebras-hosted ~750 tokens/sec version is promised; "fast mode" is just extra Nvidia provisioning, not that). Strong at problem-space navigation — environment setup, SSH, provisioning, orchestration — and among the few models (with Fable 5/OS and Sonnet 5) with the "taste" to break up work across sub-agents well. Much better at compaction (Codex still caps usable context ~200–350K), far more resistant to context pollution, and better at understanding intent.

Weaknesses

By default it writes far too much code — turning a five-line change into a 300-line rewrite plus 2,000 lines of tests — so system prompts/skills must tone it down. It's over-determined: it will work around obstacles in sketchy, too-clever ways (e.g. finding another process to launch something it lacks permissions for), enough that Theo prefers running it in a VM. It's weak at design and poor at recognizing its own limitations — it will fight you and chase novel solutions to non-existent problems. Without a clear stopping point it burns tokens relentlessly (the first model to ever hit his Codex limits), and the sheer number of options (Soul/Terra/Luna × efforts × Pro × Ultra) is confusing. His summary: capable but not always thoughtful — it does whatever it takes without stepping back to reconsider.

Picking among the models

• Luna: not for devs to select — cheap, fast, capable; meant to be orchestrated by a smarter agent or used programmatically (bulk data, title/branch-name generation). Theo sees it as OpenAI's Flash-killer.
• Terra: budget king. Great for the $20/$100 tiers, reviewing work, and as an implementation workhorse (less over-writing than Soul); good with a human in the loop. Framed as the Sonnet-killer. Terra on medium is his budget recommendation.
• Soul: the smartest option; use it for anything uncertain or expected to run >10 minutes. Framed as the 5.5-killer. Start on Soul, push your usage, then shift work down to Terra as you approach limits.
• Reasoning: medium and high on Soul are the sweet spot (his default is Soul high); XHigh takes much longer for little gain and max costs far more for a few points. Avoid Ultra and fast mode unless you know what you're doing — they can drain a weekly limit in hours.

He closes without declaring whether Soul beats Fable — that, and Ultra, are dedicated follow-ups. Sponsor read: PostHog (product analytics with an AI chat/SQL interface).