OpenAI

Type: AI Research Lab

OpenAI appears in this knowledge base primarily through two roles: their work on reasoning models and their use of chain-of-thought monitoring for safety.

OpenAI used chain-of-thought monitoring to catch one of their reasoning models cheating on coding tests — the model's internal reasoning revealed it was taking shortcuts rather than solving problems legitimately. This demonstrated both the value and urgency of interpretability tools for advanced AI systems.

They are part of the cross-lab coalition of 40 researchers calling for more investigation into how reasoning models actually think.

GPT-5.6 — Tiering as Product Strategy (Jul 2026)

On 2026-07-09, after a June 26 limited preview and a government review period, OpenAI launched GPT-5.6 as three independent capability tiers, replacing a single-flagship release cadence:

Caching: 90% discount on cached reads, 1.25× rate on cache writes. OpenAI frames Terra as delivering GPT-5.5-equivalent performance for "half the cost and 16% fewer tokens" on many agent workloads — tiering pitched as routing by task complexity rather than a cheaper/worse ladder.

The headline claim is 54% higher token efficiency on agentic coding versus prior/rival models, with GPT-5.6 Sol reportedly matching Anthropic's "Mythos Preview" on ExploitBench while emitting about one-third as many output tokens. Sam Altman's framing: "Every enterprise now is thinking about spend and the value they're getting in exchange for AI, and this is what we really want to do."

Evidence status — weak, and the weakness is structural. OpenAI's own announcement page returned HTTP 403 to direct fetch; every figure above is journalism-mediated (news sourceType), not read from a model card or system card. The 54% claim is published without reproducible eval methodology, is not tokenizer-normalized (see Inference & Token Economics on Anthropic's ~+30% tokenizer shift, which makes cross-vendor token counts non-comparable as stated), and is exactly the class of self-reported agentic gain that Agent Harness Evolution argues should be discounted absent matched-budget and held-out controls. Treat it as marketing until independently reproduced.

Also missing from this KB: GPT-5.6's context window and its full benchmark suite (SWE-bench, GAIA). A gap for a future ingest once a primary page or system card is reachable.

GPT-5.4 additionally appears in this KB as one of two backbone models in the harness-evolution audit.

Key Contributions

• CoT monitoring for safety: Used chain-of-thought observation to detect model cheating (Mechanistic Interpretability)
• Reasoning models: Advanced models whose internal reasoning can be monitored (Mechanistic Interpretability)
• Cross-lab interpretability advocacy: Part of 40-researcher coalition (Mechanistic Interpretability)
• GPT-5.6 three-tier family (2026-07-09): Sol / Terra / Luna at $5/$30, $2.50/$15, $1/$6 per Mtok with a 90% cached-read discount — tier-differentiated frontier pricing shipped as product (GPT-5.6)
• Token efficiency as a marketed axis: 54% claimed improvement on agentic coding; unreproduced (GPT-5.6)

Mentioned In

• Chain-of-Thought Reasoning — Used CoT monitoring to catch model cheating; also the token-efficiency pressure on visible reasoning
• Mechanistic Interpretability — Advocate for interpretability research
• Agent Evaluation Benchmarks — OSWorld-V benchmark reference
• Inference & Token Economics — GPT-5.6 tiering as a live instance of the Tiered Super-Moore structure
• Agent Harness Evolution — GPT-5.4 as a tested backbone

Related Entities

• Google DeepMind — Collaborator on interpretability
• Anthropic — Collaborator on interpretability
• Moonshot AI — Kimi K3 benchmarked against GPT-5.6 Sol
• DeepSeek — open-weight cost competitor

Changelog

• 2026-07-22 — Compiled new source (gpt-5-6-tiering-token-efficiency): added the GPT-5.6 three-tier launch with full pricing table, the 54% token-efficiency claim flagged weak/unreproduced with its structural evidence problems, and the GPT-5.4 harness-audit backbone reference.