1969 — [Paper] Strassen publishes efficient matrix multiplication algorithm — the benchmark AlphaEvolve would beat 57 years later. (Context: Evolutionary Algorithm Discovery)
2021–2022
Dec 2021 — [Technical Report] Elhage et al. publish "A Mathematical Framework for Transformer Circuits" — establishing rigorous methods for reverse-engineering transformers, identifying induction heads and skip-trigrams. (Anthropic Circuit Tracing)
2022 — [Technical Report] Olsson et al. discover induction heads as the mechanistic basis for in-context learning, demonstrating phase transitions during training. (Anthropic Circuit Tracing)
2022 — [Technical Report] Elhage et al. publish "Toy Models of Superposition" — demonstrating polysemanticity and why interpretability is fundamentally hard. (Anthropic Circuit Tracing)
2019–2023
2019–2025 — [Paper] Period of rapid benchmark proliferation — ~60 benchmarks developed across 8 evaluation domains for LLM agents. (From LLM Reasoning to Autonomous Agents)
2023 — [Technical Report] Bricken et al. publish "Towards Monosemanticity" — applying sparse autoencoders to extract interpretable features from transformer activations. (Anthropic Circuit Tracing)
2024
2024 — [News] Individual feature identification achieved in mechanistic interpretability — researchers can now identify specific concepts learned by AI models. (Mechanistic Interpretability)
2025
February
Feb 17 — [Paper] Xu et al. publish "A-MEM: Agentic Memory for LLM Agents" — Zettelkasten-inspired memory with dynamic note construction and linking. Later accepted at NeurIPS 2025. (A-MEM)
March
Mar 27 — [Technical Report] Ameisen et al. publish "Circuit Tracing: Revealing Computational Graphs" — attribution graphs tracing step-by-step computational paths in language models, tools open-sourced. (Anthropic Circuit Tracing)
January
May
May 14 — [Technical Report] Google DeepMind unveils AlphaEvolve — evolutionary coding agent pairing Gemini with automated evaluators. Deployed inside Google infrastructure. (AlphaEvolve)
June
September
October
November
Mid-2025
2025 — [Technical Report] AlphaEvolve achieves production impact: 0.7% global compute recovery, 23% kernel speedup, 32.5% FlashAttention improvement. (AlphaEvolve)
2025 — [Technical Report] Templeton et al. scale sparse autoencoders to Claude 3 Sonnet — millions of interpretable features, Golden Gate Bridge feature discovered. (Anthropic Circuit Tracing)
2025–2026 — [News] Mechanistic interpretability advances from individual feature identification to tracing complete reasoning paths — prompt to response. (Mechanistic Interpretability)
December
Dec 18, 2025 — [Technical Report] UK AI Security Institute publishes its first Frontier AI Trends Report — longitudinal capability assessment of 30+ models (2022–2025): cyber-task time horizons doubling ~8mo (accelerating to ~4.7mo), RepliBench self-replication <5%→>60%, frontier models surpassing the biology-PhD baseline, and universal jailbreaks present in every tested system. (AISI Frontier AI Trends Report)
2026
January
Jan 12 — [News] MIT Technology Review names mechanistic interpretability one of "10 Breakthrough Technologies 2026." (Mechanistic Interpretability)
Jan 12 — [Paper] Cheng et al. (DeepSeek/PKU) publish "Conditional Memory via Scalable Lookup" — introduces Engram (O(1) lookup) as a sparsity axis orthogonal to MoE, with a U-shaped scaling law; scaled to 27B params. (Engram / Conditional Memory)
Jan 18 — [Paper] Wei et al. publish "Agentic Reasoning for Large Language Models" — three-layer framework (foundational, self-evolving, multi-agent) organizing the field. (Agentic Reasoning for LLMs)
February
Feb 7 — [Paper] Kanagala publishes agentic AI security and red-teaming framework — threat taxonomy covering permission escalation, hallucination-driven actions, orchestration flaws, memory manipulation, supply chain attacks. (Agentic AI Security)
Feb 9 — [Analysis] Zylos Research documents 6 alignment failure modes (reward hacking, sycophancy, annotator drift, alignment mirages, rare-event blindness, optimization overhang) and formulates the Alignment Trilemma. (AI Safety 2026)
Feb 11 — [Paper] Jeddi et al. publish "LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation" (ICLR 2026) — budget-conditioned in-weights deliberation with elastic loop depth; latent reasoning without a token trace. (LoopFormer)
March
Mar 6 — [Paper] Ferrag et al. publish comprehensive review consolidating ~60 benchmarks into unified taxonomy with agent collaboration protocols (ACP, MCP, A2A). (From LLM Reasoning to Autonomous Agents)
Mar 8 — [Paper] Du publishes agent memory survey — formalizes write-manage-read loop, identifies 5 mechanism families, three-dimensional taxonomy spanning temporal scope, representational substrate, and control policy. (Agent Memory Survey)
Mar 30 — [Paper] Du publishes "Tiered Super-Moore's Law" — first systematic empirical analysis of LLM token pricing (3,237 models via Epoch AI + 318 via OpenRouter + 62 milestones, 2020–2026). Documents ~600× price decline; economy-tier half-life 1.10yr / mid-tier 1.55yr vs. the 2-year Moore benchmark; ~31.5× reasoning premium; May-2024 structural break (F = 5.74, p = 0.005); cost decline attributed ~103.7% to software/architecture and ~−0.9% to GPU hardware. (Tiered Super-Moore)
April
Apr 1 — [Paper] Hu et al. publish "Agentic Tool Use in LLMs" — unified evolutionary framework across three paradigms (prompting, SFT, RL). (Agentic Tool Use in LLMs)
Apr 2026 — [Technical Report] AlphaEvolve upgraded to Gemini 2.5 Pro for semantic evolution — now rewrites logic/control flows, not just hyperparameters. (AlphaEvolve)
Apr 2026 — [Technical Report] AlphaEvolve Service API available via Google Cloud Early Access; OpenEvolve open-source released. (AlphaEvolve)
Apr 2026 — [News] 40 researchers from OpenAI, Google DeepMind, Meta, and Anthropic call for more investigation into reasoning models' chain-of-thought processes. (Mechanistic Interpretability)
Apr 8 — [Technical Report] Meta Muse Spark — natively multimodal (text/image/voice in single transformer) with Contemplating mode that orchestrates parallel sub-agents for deeper reasoning without latency. (Muse Spark)
Apr 15 — [Technical Report] Google Gemini 3 released — claimed best-in-world multimodal understanding and most powerful agentic model. Improved tool-use, planning, native rich output. (Gemini 3)
Apr 16 — [Technical Report] Anthropic Claude Opus 4.7 released — coding-precision + safety leadership ("Project Glasswing"); leads the hardest agentic-coding splits (SWE-bench Pro ~64%). Opens the ~8-day compressed frontier window. (DeepSeek V4 — wave context)
Apr 23 — [Preprint] arXiv 2604.15726: "LLM Reasoning Is Latent, Not the Chain of Thought". Argues CoT is post-hoc rationalization, not actual computation. Implications for interpretability and CoT-based safety auditing. (Latent reasoning)
Apr 24 — [News] DeepSeek V4 released — open-weight (MIT) MoE family (V4-Pro 1.6T/49B-active, V4-Flash 284B/13B-active, 1M context); near-frontier capability (SWE-bench Verified ~80.6%, GPQA ~90–92%) at ~1/6th the cost of Opus 4.7 / GPT-5.5. Closes the compressed frontier window. (DeepSeek V4)
May
May 6 — [Preprint] OSAQ published — training-free, closed-form post-training weight quantization exploiting a stable low-rank null space in the Hessian to absorb outliers with zero inference overhead. Claims >40% lower perplexity than vanilla GPTQ at 2-bit when integrated with it. (OSAQ)
May 7 — [Paper] "Stop Comparing LLM Agents Without Disclosing the Harness" (Zhang et al.) published — formalizes the Binding Constraint Thesis and measures harness-induced variance at 7.80× model-induced variance (18.48 vs 2.37 pp², 6/9 ranking reversals) in a controlled 3×3 SWE-bench Verified factorial; proposes the ETCSOVG Harness Card disclosure standard. The "edge is the harness, not the model" thesis, measured. (Stop Comparing)
May 17 — [Preprint] VeriCache published (Yao et al.) — turns lossy KV-cache compression lossless by drafting from the compressed KV (GPU HBM) and verifying against the full KV (CPU/remote) in parallel. Up to ~4.3× throughput, output identical to full-KV (KL < 0.01 nats), ~25–40 accepted tokens/round; composes with EAGLE/MTP to 4.35×. Opens the KV-cache lane in this KB. (VeriCache)
May 27 — [Paper] Harness-Bench published (Yao et al.) — diagnostic benchmark (106 sandboxed tasks, 5,194 execution trajectories) isolating configuration-level harness effects across model backends; names execution-alignment decoupling as the dominant failure class and argues performance must be attributed to model-harness configurations, not models. (Harness-Bench)
June
Jun 9 — [Technical Report] Claude Fable 5 and Claude Mythos 5 launch — 1M-token context, 128k max output, always-on adaptive thinking, and a new tokenizer emitting ~30% more tokens for the same text. Fable 5 runs safety classifiers that can return stop_reason: "refusal" with no billing on zero-output refusals. (Anthropic Release Notes)
Jun 16 — [Paper] "How Inference Compute Shapes Frontier LLM Evaluation" (McFadyen et al.) published — across up to 12 frontier models × 7 benchmarks (SWE, math, medicine, cyber) under controlled inference-scaling interventions, fixed single-budget scores increasingly understate newer models, which have a higher ceiling that only opens at generous budgets. Recommends reporting capability as a function of inference-time compute. Makes inference budget a first-class evaluation axis. (How Inference Compute Shapes Evaluation)
Jun 25 — [Technical Report] Anthropic deprecates fast mode for Claude Opus 4.7, removal scheduled 2026-07-24. (Anthropic Release Notes)
Jun 26 — [Technical Report] Anthropic consolidates rate limits — Sonnet and Haiku limits now match Opus at every tier; usage tiers collapsed to three (Start, Build, Scale). (Anthropic Release Notes)
Jun 26 — [News] OpenAI opens a limited preview of GPT-5.6, ahead of a government review period. (GPT-5.6)
Jun 30 — [Technical Report] Claude Sonnet 5 launched — $2/$10 per Mtok, launched as introductory through 2026-08-31 with $3/$15 to follow. The step-up was cancelled on 2026-08-10 and $2/$10 became the standard price (see Aug 10 below). 1M-token context, 128k max output. Adaptive thinking on by default with manual extended-thinking budgets removed (400 on attempt); non-default sampling params now 400; new tokenizer bills ~30% more tokens for identical text. (Anthropic Release Notes)
July
Jul 1 — [Technical Report] Anthropic restores access to Claude Fable 5 and Claude Mythos 5 following a prior suspension. (Anthropic Release Notes)
Jul 2 — [Technical Report] Anthropic ships the agent-memory-2026-07-22 beta header, replacing managed-agents-2026-04-01 on memory-store endpoints — server-defined stable ordering, depth restricted to 0/1/omitted, segment-exact path_prefix required to end in /, prior page cursors invalidated. Agent memory becomes versioned platform surface. (Anthropic Release Notes)
Jul 2 — [Preprint] Spec-AUF published — truncates block-drafter cross-entropy support at the drafter's first predicted failure, concentrating supervision on the accepted prefix. τ 2.40→2.61 on Qwen3-8B (DFlash, mean of six benchmarks); transfers to Domino (2.56→2.68). No inference-pipeline change. (Spec-AUF)
Jul 5 — [Preprint] "Quantize the Target, Quantize the Drafter" published — quantization-aware distillation of the target plus a two-stage-trained block-diffusion drafter on a single NVIDIA A10G. 6.978× average speedup meeting required quality thresholds; 3rd overall, Efficient Qwen Competition (ICML 2026 Workshop on AdaptFM). Direct evidence quantization and speculative decoding compound (bounded — see the flagged arXiv:2505.22179 conflict). (Quantize the Target)
Jul 5 — [Preprint] "Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents" (Yi & Song) published — the harness changes an agent's multi-step beliefs (progress, risk, recoverability, next action, …) even at matched terminal success; introduces a belief-rollout diagnostic (arrival + growth decomposition) and BIWM. Outcome parity is not harness equivalence. Freshest of the July harness cluster. (Belief Divergence)
Jul 6 — [Preprint] DSpark published (DeepSeek author roster incl. Wenfeng Liang, Damai Dai) — semi-autoregressive drafting plus confidence-scheduled verification. 60–85% faster per-user generation at matched throughput vs. the MTP-1 production baseline, measured under live user traffic in the DeepSeek-V4 serving system; "shifts the Pareto frontier of our serving system." (DSpark)
Jul 8 — [Technical Report] Anthropic adds API key expiration — preset/custom/"Never" durations, pre-expiry email reminders, expires_at on Admin API key listings. (Anthropic Release Notes)
Jul 9 — [News] OpenAI launches GPT-5.6 as three tiers — Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per Mtok, with a 90% cached-read discount and 1.25× cache-write rate. Headline claim: 54% more token-efficient on agentic coding; Sol reportedly matches Anthropic's "Mythos Preview" on ExploitBench at ~one-third the output tokens. Vendor claim, unreproduced. (GPT-5.6)
Jul 10 — [Technical Report] Anthropic's Dreams managed-agent research preview expands to Claude Fable 5 and Claude Sonnet 5. (Anthropic Release Notes)
Jul 14 — [Paper] Wang et al. publish "Rethinking the Evaluation of Harness Evolution for Agents" — automatic harness evolution does not consistently outperform simple test-time scaling under matched feedback/inference budget, and generalizes poorly to held-out tasks. Terminal-Bench 2.1, GPT-5.4 and Claude Opus 4.6. Establishes matched-budget + held-out evaluation as the methodological floor for agentic claims. (Rethinking Harness Evolution)
Jul 14 — [Technical Report] Anthropic ships Admin API user management (beta) to all Claude Enterprise orgs — member lookup, role changes, removal, invites, groups, custom roles. (Anthropic Release Notes)
Jul 15 — [Technical Report] Anthropic Claude Platform release notes current through this date — the first-party changelog ingested as this KB's highest-authority July source. (Anthropic Release Notes)
Jul 16 — [Analysis] Moonshot AI releases Kimi K3 — ~2.8T-parameter MoE ("LatentMoE", 896 experts / 16 active per token, ~1.8% activation), 1M context, native vision; described as the largest open-weight model to date. API live at $3/$15 per Mtok — up from K2.6's $0.95/$4, the most expensive Chinese-lab model to date. Tops Arena.ai's Frontend Code leaderboard ahead of Claude Fable 5 while Moonshot's own eval places it behind Fable 5 overall. Open weights promised 2026-07-27. (Kimi K3)
August
Aug 10 — [Technical Report] Anthropic cancels the Claude Sonnet 5 price step-up — the $2/$10 per Mtok rate, announced at launch as introductory through 2026-08-31 with $3/$15 to follow, becomes the standard price. The scheduled increase does not occur. Prior-gen Sonnet 4.6 remains $3/$15, so the permanent rate sits a third below the incumbent it replaced. Verified 2026-08-13 against the pricing page and the 2026-08-10 release-notes entry. Note the tokenizer offset: Sonnet 5 bills ~30% more tokens for identical text, so per unit of text the cut against Sonnet 4.6 is ~13%, not 33%.
Aug 12 — [Technical Report] xAI releases Grok 4.6 — API list price $2.00 / $6.00 per Mtok, with a "fast variant which is twice the price"; xAI's own announcement states no change against Grok 4.5. That puts xAI on the same input rung as Claude Sonnet 5 and GPT-5.6 Terra ($2.00) while pricing output at $6.00 — 40% below Sonnet 5's $10 and half GPT-5.6 Terra's $12. Self-reported Artificial Analysis Intelligence Index 61, GDPVal-AA v2 1753, Terminal-Bench v3.0 26%. Verified 2026-08-18 against xAI's own announcement. Recorded into the Token Price Index the same day — the first xAI row it has ever carried, closing a provider-shaped hole in the rung evidence.
Aug 13 — [Technical Report] DeepSeek V4-Pro reaches general availability — the DeepSeek-V4-Pro-0813 checkpoint ships to app, web and API, taking the 1.6T-total / 49B-active MoE flagship out of preview. Open-weight (MIT) and 1M-context per the V4 family recorded at Apr 24 above; the licence is carried forward from that entry, not re-verified primary today. Verified 2026-08-18 against DeepSeek's own changelog: "The GA release of DeepSeek-V4-Pro has been rolled out on the APP, Web, and API." Shipped alongside the deepseek-harness agent runtime, recorded on the harnesses rung rather than here.
Aug 13 — [Technical Report] Google ships Gemini 3.7 Flash GA with a published expiry date on its price — introductory $0.75 / $3.75 per Mtok (cached input $0.075), stated on Google's own pricing page as "$0.75 through December 31, 2026. $1.50 starting January 1, 2027." and "$3.75 through December 31, 2026. $7.50 starting January 1, 2027." Every element doubles on 1 Jan 2027, cache storage included ($0.50 → $1.00 per Mtok-hour). Google applied the identical dated rate to the incumbent Gemini 3.6 Flash, which the Token Price Index recorded at a flat $1.50/$7.50 on 2026-08-03 — a 50% cut to the incumbent and one shared expiry across the whole live Flash generation. Prior-generation Gemini 3.5 Flash is untouched at a flat $1.50/$9.00, so the promotion is generation-scoped, not line-wide. Verified 2026-08-18 against the Gemini API pricing page. (The date Google cut 3.6 Flash is not established — the Index's prior observation is 2026-08-03 and the next is today, so the change is bounded to that window, not dated.)
Aug 14 — [Technical Report] Alibaba releases Qwen3.8-27B — a 27B dense native vision-language model (text, image and video in, text out) under Apache 2.0, 262,144-token native context extensible to ~1M via YaRN, weights on Hugging Face and ModelScope. Scores 52 on the Artificial Analysis Intelligence Index (measured 2026-08-17), ranked #1 of 135 among open-weight models of comparable size. Verified 2026-08-18 against the model card and Artificial Analysis. No API list price is recorded: the model is weights-only here, so it does not enter the Token Price Index until a hosted rate is verified against a primary source.
Aug 16 — [Technical Report] DeepSeek's price rise takes effect — the first outright increase by the lab whose public identity in this KB is cost disruption. Warned on 6 Aug with no rates attached; effective 16:00 UTC on 2026-08-16 per DeepSeek's own changelog: "we will adopt peak/off-peak pricing, with off-peak prices set at half of the peak-hour prices." Flat pricing is replaced by a two-tier structure (peak = 01:00–04:00 and 06:00–10:00 UTC; all other hours off-peak). V4-Pro output moves from a flat $0.87 to $1.98 off-peak / $3.96 peak (+128% / +355%); V4-Pro cache-miss input from $0.435 to $0.66 / $1.32 (+52% / +203%); V4-Pro cached input from $0.003625 to $0.022 / $0.044 — a 6.1× to 12.1× rise and by far the largest single move, which falls hardest on exactly the cache-heavy agentic workloads the harness layer generates. V4-Flash output moves from $0.28 to $0.66 / $1.32 (+136% / +371%). Verified 2026-08-18 against the pricing page and changelog; prior rates from the Token Price Index rows observed 2026-08-03. The peak window is unchanged from what the Index recorded on 2026-08-03 (09:00–12:00 and 14:00–18:00 Beijing = 01:00–04:00 and 06:00–10:00 UTC); what the Index did not anticipate is that the base level would rise underneath it.