The posts under <a href="/hypotheses/llm-ceiling/">The Ceiling</a> test whether LLMs are running into capability limits. This page draws the operational inference: if you're shipping with LLMs, when do you commit, when do you wait, and what's worth doing regardless? ← Hypothesis tracker
Build Now
Wait
Match your build horizon to the release cadence
The most overlooked variable in the build-vs-wait decision is the rate at which the frontier ships. Anthropic released Claude Opus 4.8 on May 28 2026 — forty-one days after Opus 4.7 landed on April 16, with pricing unchanged at $5 per million input / $25 per million output, a Fast Mode running at 2.5× speed for 3× lower cost, and a $65 billion capital raise at a $965 billion valuation on the same day. The cycle from 4.6 to 4.7 had taken roughly seven months. The cycle from 4.7 to 4.8 was 41 days. Three Opus releases have landed in three calendar months, with a Mythos-class model — independently expected to deliver SWE-Bench Verified scores in the low-90s — held back on cybersecurity rather than capability and being prepared for general availability in the next several weeks. The competitive cadence is similar: OpenAI shipped GPT-5.5 in April and Codex-class agents have continued to land monthly; DeepSeek went from V3.2 to V4 to V4 Pro Max in a single quarter; Gemini 3.1 Pro updates have been monthly. The base case for the next twelve months is that whichever capability gap your product is solving with a workaround today will be partially or fully closed by a frontier release in the calendar interval between deciding to build the workaround and actually shipping it.
The operational rule that follows is to size your workaround projects against the frontier release cadence rather than against your own product roadmap. A workaround that takes six weeks to build and ship is acceptable only if you have a specific reason to believe the next two frontier releases will not narrow the capability gap it covers. A workaround that takes six months is, by default, building against capability that will exist by the time you ship. Three concrete heuristics fall out: never build a multi-month bridge for a capability that has visibly improved on independent benchmarks in each of the last two quarters — wait one more release cycle and re-evaluate. Always build a multi-week bridge against a capability that has plateaued for two or more consecutive frontier releases in the same family (Terminal-Bench movement at the frontier has been narrow across multiple releases, for example, so terminal-agent reliability infrastructure is a build-now bet, while broad coding-agent capability is a wait bet). And track the price-per-token curve of the model you depend on — Anthropic just cut the Opus Fast Mode price 3× without changing capability tier, which means your unit economics on inference-heavy workflows are improving faster than your time-to-build can capture. When cadence and price are both moving against the workaround, the optimal posture is to ship a thin integration on current capability and rebuild it when the next release lands; the rebuild cost is almost always less than the build-against-fixed-capability cost.
Evidence: The 41-Day CycleHedge
Bet on the axis, not the ceiling
Capability is not a single number, and the ceiling debate is being argued on two axes that have decoupled. On the narrow axis — where verification is mechanical (proof checking, test-suite-validated code, formal logic, structured data tasks) — peer-verified autonomous-discovery results from frontier labs have been arriving every three to six months for the past eighteen: AlphaProof IMO silver in 2024 with its methodology in Nature in November 2025; DeepMind Deep Think and OpenAI both at IMO gold in 2025; AlphaEvolve improving bounds across 67 problems in November 2025; the OpenAI Erdős unit-distance proof on May 20 2026, peer-verified by a nine-author panel that includes the researcher whose explicit job is policing AI math claims against the published literature. The pace is steady. On the broad axis — where capability is measured by independent multi-task benchmarks — the picture is mixed: SWE-Bench Verified frontier scores moved from roughly 65% in Q1 2025 to 88.7% vendor (82–83% independent on vals.ai) by April 2026 with GPT-5.5, with Claude Mythos preview now near 94%, and SWE-Bench Pro as the contamination-resistant successor sitting in the 55–65% range with substantial headroom; LMArena's top tier has held inside the 1,450–1,561 Elo band over the last six months but the Arena Expert methodology was sharpened in November 2025, which compresses cluster scores mechanically. The honest read: progress on both axes is ongoing in the recent window, with no clean inflection visible against the 12-month frame the ceiling thesis is built around.
The operational mistake is betting your product or investment on a single axis. A team building an autonomous-discovery system in a domain with mechanical verification — drug screening, formal verification, theorem proving, anything where output correctness is checkable — is exposed to the narrow axis and should plan for continued capability gains over the next 12 months at roughly the same release cadence as the past 18; the workaround is likely to become cheaper than the build before the year is out. A team building a general assistant or agent in an unstructured environment is exposed to the broad axis and should plan for capability gains to come from a combination of cost compression, tool integration, and continued one-to-two-percentage-point monthly improvement on contamination-cleaner successor benchmarks (SWE-Bench Pro, ARC-AGI-2) rather than from leaderboard-topping frontier-model releases; the broad axis is moving but not leaping, so building around current capability with a quarterly upgrade rhythm is the right default. The hedge is to instrument both axes against your specific task on a three-to-six-month comparison window, not on multi-year averages that wash out the inflection. Track at least one independent narrow benchmark relevant to your domain and at least one independent broad benchmark monthly; the moment your task starts moving meaningfully on one axis without the other catching up, you know which axis your product economics depend on, and the strategic bet writes itself.
Evidence: The Proof That Didn't Move the ScoreLast updated: 2026-06-01 · hypothesis tracker