The Scientist

OpenAI staff confirm GPT-5.5 and GPT-5.6 effort settings are not directly comparable. The model picker is gone and Luna/Terra/Sol routing is opaque. Every regression suite, prompt contract, or effort×cost curve calibrated on 5.5 is now measuring a moving target. 'Start lower than on 5.5' is community folk wisdom filling in for the calibration data OpenAI didn't publish.

The tiers hide a real compute split. Max is one model thinking longer. Ultra parallelizes across subagents, which is a different latency and token-budget profile. The live hazard is spawn_agent: with no model or effort selection, Sol Ultra recursively spawns Sol Ultra by default. That is a cost blowout waiting for a production quota.

The cheapest experiment is the Raschka hypothesis (459K views): Luna at higher effort matches or beats Terra at lower effort for less money. That is consistent with OpenAI's claim that Luna at lowest effort beats GPT-5.5 at highest effort at 25x lower cost on health tasks. The thing that claim doesn't tell you is whether it holds on your distribution. If it does, a one-day paired eval yields a material spend cut.

Then Statsig: the vendor whose 36 variants you must benchmark now owns the platform holding many teams' benchmark records, meaning experiment definitions, exposure logs, and metric outcomes. Nothing suggests misuse, but the conflict-of-interest geometry is new, and the contractual status of your experiment data just changed hands. Confirm export paths while acquisition terms are still being socialized.

Against an unpinned GPT-5.6 endpoint you aren't running an experiment. You're sampling a routing distribution you can't observe.

The headline number is per-token price. That's not the decision. Grok 4.5 is cheaper per token than Opus 4.8, but it also reportedly burns 4.2x fewer output tokens on identical SWE-Bench Pro tasks, which compounds into roughly an order-of-magnitude per-task cost gap. The thing that number doesn't tell you is failure rate. Grok 4.5 scores 53% on DeepSWE 1.1 vs Fable 5's 70%, so on hard tasks the savings leak back out as retries. Price times token burn times failure rate is the actual cost function, and only your harness computes it on your traffic.

Meta's entry is the structural event. Muse Spark 1.1 is its first paid model — $20 free credits, Zuckerberg pledging 'very aggressive' pricing — a pivot from open-source Llama to proprietary access, under Wall Street pressure to monetize AI infra without a cloud business. The claimed wins over Gemini 3.1, Opus 4.8, and GPT-5.5 on tool use and financial analysis ship with zero named benchmarks, sample sizes, or ablations. Prices are verifiable. Quality claims are keynote copy until someone runs them. The free credits are budget for that falsification.

All three incumbents are converging on manager-worker orchestration. Anthropic's Managed Agents hold 92% of SWE-Bench Pro capability at 63% cost, and 96% of web-automation benchmarks at <50%, via sub-agent cache isolation. OpenAI's Programmatic Tool Calling runs 4 parallel agents at a third fewer tokens. A cheap executor under an expensive planner works with any model pair. That makes it a spike, not a platform bet.

Background pressure worth pricing in: US developer traffic to Chinese models has passed 40%, with DeepSeek V4 undercutting proprietary APIs. That's real contract-negotiation leverage, subject to the usual sovereignty caveats.

Meta selling model access means free frontier weights persist by vendor choice, not by law. Anything you depend on is now revocable, so treat availability as a variable and archive accordingly.

The mechanism: restored Claude Fable 5 silently routes some cybersecurity and coding queries to the weaker Opus 4.8 — no version bump, API header, or changelog. Users report degraded biology and coding performance post-restoration. Anthropic argues any capable model could produce the same exploits, meaning the guardrails are a compliance artifact, not a capability boundary — but your prompts don't care. Production behavior changed and nothing in your stack was notified.

The precedent stack makes this systemic: the Fable 5 suspension was the first-ever US export-control directive against a deployed model; OpenAI staggered the GPT-5.6 rollout at government request; both Anthropic and OpenAI delayed their newest models after weeks of negotiation with the administration. 'Next-gen model on date X' is now a distribution with a fat regulatory tail.

Vendor trust is the third leg: an OpenAI privacy engineer testified the company misled courts for two years about its ability to search ChatGPT logs — it had run searches on large anonymized samples pre-litigation. Any compliance posture leaning on vendor log-handling representations is now contradicted by the vendor's own sworn testimony.

Three assumptions broke at once: version implies behavior (guardrail routing), availability implies schedule (export directives), vendor representations imply controls (log testimony). Each has a cheap answer — drift canaries, fallbacks, upstream redaction — and none can be retrofitted after the incident.

When a guardrail can swap your model mid-conversation, the only version number you can trust is the one your drift canary prints every morning.

DSpark's most decision-relevant number is the easiest to misuse: V4-Flash throughput gains run +51% at 80 tok/s/user but +661% at 120 tok/s/user. That is a batching-regime effect, not a flat multiplier. Lift either figure into a capacity plan without benchmarking at your own SLO and the forecast is fiction.

Mechanically, DSpark bolts a Markov-head parallel drafter onto frozen weights. Per-user generation runs 57-85% faster than the prior MTP-1 drafter with zero accuracy loss, accepted tokens per verification round beat EAGLE-3 by 26.7-30.9% (DFlash by 16-18%), and added latency per round is just 0.2-1.3%. The gains held on Gemma4-12B, a separate lineage from Qwen, so the design looks architecture-general and portable to an EAGLE-3/MTP-style stack. V4-Flash and V4-Pro checkpoints are on Hugging Face today under MIT.

It lands in a convergent efficiency wave. Unsloth's NVFP4 quants claim 2.5x faster inference and fit a 27B model on 24GB VRAM; Qwen3.6 35B-A3B posts 17,561 tok/s on a B200 (65 tok/s at 128k context on dual B60); Cohere open-sourced hardware-aware, batch-adaptive speculative decoding in vLLM; a Gemma challenge hit 315 TPS lossless on a single A10G; PrismML compressed a 27B Qwen 3.6 onto an iPhone Pro. All figures are vendor- or challenge-reported on specific hardware. The thing these numbers don't tell you is how they compose on your traffic. Stacked, though, a 27B-class model on one consumer card with a modern drafter is a credible production tier, and that changes the self-host math before the next inference commit renews.

A speculative-decoding speedup is a function of your target tokens-per-second, not a property of the model — benchmark at your SLO or you're quoting someone else's regime.