Reasoning Models & the AGI Debate

An explainer that reads the debate over AI-produced math results against the history of computer proofs. The four color theorem in 1976 was proved with computer calculations no person could follow to the end, which cause

Company-reported 87% on the embodied reasoning benchmark ERQA. Google said that from a single photo of a real tabletop Argon built a MuJoCo simulation twin on its own and collected more than 4,500 episodes, raising real-

The GPT-6 generation reached the default chat of ChatGPT, which has more than 1.2 billion weekly users. Because it begins answering while it keeps thinking, GPT-6 Instant starts answering 44% sooner on average than GPT-5

The company said requests up to 100,000 tokens cost 90% less than Haiku 4.5 and longer requests 50% less, for about 75% lower average running cost. The same day it halved Sonnet 5.5 cache-read prices and introduced a mon

It also published 10 summaries of the model's reasoning, compute estimates expressed as ChatGPT Pro usage and statistics on attempted problems, saying the average result used about 3 hours of ChatGPT Pro thinking. The fi

Company-reported scores: DeepSWE v1.1 77.9%, AutomationBench 51.3% (first), LVBench 91.7%, CWE-bench v1 68% (tied first). The introductory price is $2/$10 per 1 million tokens (later $4/$20), and Google said it is taking

A mid-tier update released alongside Anthropic's Sonnet 5.5 in the same week. That two labs simultaneously touted "nearly the same work at 1/5 of the frontier price" shows the competitive axis of the second half of 2026

The vendor's claim itself — that the mid tier has climbed to within two points of the top tier — is this week's event. As the performance gap between tiers narrows, price becomes the deciding factor, and the reason to us

The GPT-6 generation came down to lower price tiers just 3 weeks after its announcement. If the claim that they reproduce the top model's performance at around 1/10 of the cost holds, the bottleneck in deploying agents s

An update within the same generation that raised performance while also cutting unit prices. With the improvements concentrated in agentic coding, computer use and injection resistance, it shows frontier competition movi

A summary of a conversation that uses the numbers on model turnover cycles to explain why the call to slow down came now, of all times. In one model family, the gap between top models shrank from 877 days to 336 and then

A case of solving, through architecture, a chronic limitation of diffusion language models: "state breaks at chunk boundaries". A substantive advance that arrived just as the observation topic "diffusion language models"

This sums up a short report that bundles together a safety warning and a performance breakthrough from the same week. Its construction is unusual: it directly asks two commercial models whether they are confident that th

An attempt to break RSI down into stages and make it something to measure rather than a slogan. The theoretical axis of this week's simultaneous cluster of RSI and harness self-improvement papers on arXiv.

A case of releasing, down to the weights and data, a result that reached the IMO gold-medal range through pure natural-language reasoning without a formal checker — greatly widening the reproducibility of top-level mathe

According to OpenAI, about 100 agents working together for about 50 hours also disproved the regularity of the Euler equations without external forcing; they reached the result about 88 hours after the first agents were

A structure that feeds the routing logs generated during serving back in as the next training data — an early prototype of a self-improvement loop in which the boundary between deployment and training disappears.

Along with the characterization that we have moved from "AI that answers" to "AI that does the work itself," the new model's ability to operate a computer on its own and finish tasks is presented as what sets it apart. T

Since these are the frontier-lab chief scientist's own words, it is a primary source for what he said. Written just after the company declared that its GPT-6 Astra model met the Critical cyber threshold, it puts a declar

Overlaying forecasts made by executives at several AI companies onto a single timeline produces this sequence: agentification in 2026, automation of AI research itself in 2027, long research processes carried through to

A claim that an agent swarm carried out, in a short time, a large formalization project that the human community had been pursuing over a span of years. Because the output is a checkable artifact (a Lean proof), it can,

The start of the GPT-6 generation. Released the same week as Anthropic's Fable 5.1 with the same price tag of $10/$50 per 1M tokens, so frontier pricing has effectively converged, and the center of benchmark competition

Security-specific derivative models are split off as official variants of the frontier lineup — vulnerability detection and patching become an independent axis of the model family

MSL keeps to its closed-API route while putting agent efficiency (fewer tool calls and tokens) forward as its headline metric

The competitive axis for frontier models shifts from coding to autonomous scientific research. Tiered deployment that opens the top model only to screened users has hardened into a product line, and the fact that the pri

Focused on long-running agents. The safety description says only that safeguards were calibrated to its capabilities, and no separate model card is linked

A shift in the product axis from chatbots to resident autonomous agents

A community-led release of audit trails, rather than a lab announcement, plus real-time olympiad evaluation with independent grading emerges as a new practice for checking results

A new form: joint training with a coding-product company. States that competitors' figures were quoted from each company's system card

A defining 2026 remark showing a shift from the earlier, cautious way of talking about AGI timelines to a more assertive framing

An event that revealed the definition of AGI had really been a contractual and governance device

The variance explained by the first principal component across benchmarks fell from 92% in 2023–24 to 64% after reasoning models arrived — an argument for specialization beneath apparent generality

A conditional shift, citing the emergence of fluid intelligence that adapts to novelty at test time

AI-generated papers passing review entered the formal scholarly record. The debate over how to evaluate automated research got going in earnest

The axis moved from static problem solving to interactive exploration and skill acquisition without instructions

LLM skepticism capitalized as a $1 billion-scale bet on an alternative route

Became the default model for Free and Pro plans in claude.ai and Claude Cowork; Anthropic reported that early-access developers strongly preferred it to its predecessor and often to Opus 4.5 from November 2025

A logical structure in which the end of scaling and the imminence of AGI are two sides of the same claim

ARC-AGI-2 is effectively saturated. The announcement does not discuss AGI, framing the upgrade instead as accelerating science, research and engineering

Company-reported for GPT-5.2 Thinking: GDPval wins or ties 70.9% (GPT-5: 38.8%), ARC-AGI-2 (Verified) 52.9% (GPT-5.1 Thinking: 17.6%), SWE-bench Verified 80.0%

Showed that the key technical change of 2025 was iterative loops of search and checking

Open-weight reasoning models; DeepSeek claimed V3.2-Speciale reached gold-level results at IMO, IOI and other 2025 competitions

Company-reported: at medium effort it matches Sonnet 4.5’s best SWE-bench Verified score while using 76% fewer output tokens

The biggest skeptic camp splitting off organizationally

The ARC-AGI-2 SOTA doubled in a single leap

A release that made emotional and creative interaction an explicit optimization target

OpenAI said Thinking adapts its thinking time more dynamically, roughly twice as fast on the fastest tasks and twice as slow on the slowest, and that the name reflects an iterative upgrade within the GPT-5 generation

An open-weight reasoning agent trained to interleave thinking with tool calls; Moonshot claims it sustains 200–300 sequential tool calls without human intervention

Quantitative evidence of a failure to transfer to real-world knowledge work

Moves the definition of AGI from rhetoric to a psychometric scale; the biggest deficit is long-term memory storage

Anthropic said it gives coding performance similar to Sonnet 4, state of the art five months earlier, at one-third the cost and more than twice the speed; released under ASL-2 after testing found only limited CBRN risk

Anthropic said it observed the model maintaining focus for more than 30 hours on complex multi-step tasks; released under ASL-3 protections together with the Claude Agent SDK

The first frontier LLM to appear in a journal after formal peer review. A crack in the practice of relying on corporate technical reports

A strategy of giving away a low-cost coding model through Copilot, Cursor, Cline and others to win market share first

Billed by DeepSeek as its first step toward the agent era; the API's deepseek-chat and deepseek-reasoner became the non-thinking and thinking modes of one model, merging what had been separate V3 and R1 offerings

The disillusionment phase of AGI expectations; Altman himself acknowledged that, by the definition of AGI, something important is still missing

Company-reported 74.5% on SWE-bench Verified; priced the same as Opus 4, with Anthropic recommending the upgrade for all uses

That a general-purpose model achieved this end-to-end in natural language, without formalization, is the key piece of evidence in the debate over generality

An approach that raised RL compute to pretraining levels. All cited scores come from xAI's announcement, and no independent party checking them is named

Partly rebuts the evaluation design of Apple's paper, while acknowledging that some failures are genuine reasoning limits

The first open-weight reasoning model family from a European lab. A paper on the training infrastructure and RL algorithm was published at the same time

The most widely cited empirical case for skepticism: reasoning effort actually decreases past a threshold

Company-reported: SWE-bench Verified 72.5% for Opus 4 and 72.7% for Sonnet 4, Terminal-bench 43.2% for Opus 4. Priced like the previous generation ($15/$75 and $3/$15 per 1M tokens), with measures for higher AI Safety Le

A quantitative forecast that reasoning training will reach the frontier of total training compute within about a year, after which its growth rate will converge to around 4× per year

Reasoning models shifted from chains of thought to agents; the axis of competition moved from how long a model thinks to how long it can work autonomously

The flagship document of the short-timelines camp

A declaration that thinking was moving from a special mode to the default

LG's open-weight line specialized for math, science and coding reasoning; company-reported scores for the 32B were 94.5 on the 2025 CSAT math section and 90.0 on AIME 2024

The reference point for the cautiously optimistic camp

Hybrid architecture, 23 languages, specialized in RAG, grounding and tool use. A decentralized training approach using self-refinement and model merging

The full release following QwQ-32B-Preview; Alibaba claims performance comparable to DeepSeek-R1 (671B total, 37B active) at 32B

A non-reasoning model that does not think before responding; company-reported GPQA 71.4% and SWE-bench Verified 38.0% trailed o3-mini (high) at 79.7% and 61.0%, while it led on MMMLU (85.1% vs 81.1%) and SWE-Lancer Diamo

Declared the position that reasoning should be an integrated capability of frontier models, not a separate model

Split out a reasoning mode behind a Think button. The 200,000-GPU cluster in the announcement was mentioned for training the next model; the announcement does not say it was all used to train Grok 3

Previewed in December 2024, it launched the same day in ChatGPT Plus, Team and Pro and in the API for usage tiers 3–5, and free users could try it through the 'Reason' button — the first reasoning model offered to free C

R1's claim of low-cost training shook the moat argument behind AI capex

The next frontier after MMLU saturation; models at the time scored under 10%

Proved that reasoning ability is not a secret recipe of closed labs; self-checking, re-examination and strategy switching emerged. The same report also published the RLVR recipe and a recipe for transferring the capabili

The designer of ARC put capital and an organization behind the alternative route of program synthesis

The most direct AGI declaration yet by a frontier-lab CEO, moving the goal on to superintelligence

Google's first reasoning model, taking the route of showing its thinking process to users

Google claimed 2.0 Flash outperforms 1.5 Pro on key benchmarks at twice the speed, launched the Multimodal Live API and Deep Research, and presented agent prototypes such as Project Astra, Project Mariner and Jules

The commercialization of inference-scaling: users pay for the model to think longer

The first open reasoning model after o1

Named the approach of replacing the reward model with a verifier function, and released every stage of the pipeline

The first public statement of skepticism about pretraining scaling

Ignited the scaling-wall debate

The first commercial reasoning model; on AIME 2024, from 12% for GPT-4o to 74% for o1

Turned the research-automation discourse into a concrete figure: $15 per paper

Image generation started not with an in-house model but through integration with an external model (Black Forest Labs FLUX.1)

Demonstrated that test-time compute can be more efficient than scaling up parameters

Self-training grounded in a formal language reached olympiad level. DeepMind stated explicitly that it would not have been possible without Lean and Mathlib

OpenAI said it is more than 60% cheaper than GPT-3.5 Turbo; it replaced GPT-3.5 for ChatGPT Free, Plus and Team users

Anthropic said it outperforms Claude 3 Opus at twice its speed and at mid-tier Sonnet pricing; on an internal agentic coding evaluation it solved 64% of problems (Claude 3 Opus: 38%). The UK AISI ran a pre-deployment saf

Expanded the context 16× over the previous version

Opus and Sonnet available on launch day in claude.ai and the Claude API, generally available in 159 countries, with Haiku to follow. Per 1M tokens: Opus $15/$75, Sonnet $3/$15, Haiku $0.25/$1.25; assessed at ASL-2 under

An early case of formally verifiable rewards driving reasoning training. Produces human-readable proofs

Its first product. Made wit and a rebellious streak explicit design goals, and presented real-time knowledge from the X platform as its differentiator

The starting point of a new frontier lab made up of people from DeepMind, OpenAI and Google. The list of founding members follows Wikipedia; an official xAI announcement has not been checked