This question was picked from the 20 records added that week. The answer is in the records, and understanding starts with the Five-Minute AI Lessons.
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 use frontier models shifts from task difficulty to latency and cost. But jumps such as 10.3%→70.6% on Terminal-Bench have no outside check of whether they compare like with like, so they are kept as company claims only.
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 moving from top performance to price per unit of performance.
The first case of a competing lab putting numbers on what leaks out when open weights are released without safety measures. The debate in the open-weight lineage (axis ⑨) shifts from a "performance gap" to a "safeguard gap", and the issue becomes not the release of weights itself but what was not released along with them. The interests at play when a competitor evaluates a competitor should be taken into account.
What ties the events together — my own reading.
One company released its second-tier model and said it scores just 2 points below its own top model. The very next day, another company released its second tier too, saying it offers near-top intelligence at a fifth of the price. By coincidence, the two models cost exactly the same — per 1 million tokens, $2 for text going in and $10 for text coming out. But every one of these scores was measured by the company that made the model. There are no results yet from anyone else grading the same test.
On the same day, one company created a separate 'Ultrafast' tier. It makes text come out up to 8 times faster, but it's only available on expensive plans. Smarts and speed have split apart, and each now has its own price. The open-source side answered the same week too — it published real measurements showing that if you split reading the text and writing the text between different machines, the time a person waits stays within 0.05 seconds even at its worst. That same wait used to stretch to 0.26 seconds when many people piled on at once.
One company took apart another company's openly released model. Simple tricks got past its safeguards anywhere from 64% to 100% of the time, and with a bit of tinkering, the rate at which it refused dangerous requests reportedly fell from 95% to 6%. The issue isn't that the model was opened, but what wasn't opened along with it: the safeguards against misuse. Still, this is a rival grading a rival, so it shouldn't be taken at face value, and since the numbers are public, they can be challenged on the same footing.
Take the week's quiz (10 questions)
Records and notes, together or apart.
Model suppliers have begun issuing workforce credentials themselves. The diagnosis is that the bottleneck is not model performance but "people who can deploy it in companies", and if labs hold the credentials, the deploy
The point at which deep-research-style capabilities came down to an 8B open-weight model. Using citation accuracy and recall as evaluation axes matters, because in report generation the problem is not sentence quality bu
As skill and plugin marketplaces grow, the supply chain becomes an attack surface. A structure that runs vetting through benchmarks from multiple participants rather than a single vendor follows the path package reposito
A rare case of a large bank in a regulated industry disclosing seat counts and processing volumes. 16,000 people and 120,000 items a day are operating scale, not a "pilot", and the target of 50% of developers shows codin
The point at which generation paths outside autoregression matured from the claim that "parallel generation is fast" to the question of "how to preserve dependency structure". A design that tries to solve the respective
The optimization target of harness engineering has widened from "what to fix" to "what to test". Abstracting failures into patterns to compose the curriculum automatically takes over the place where people used to pick s
In a company presentation on sovereign transformation in healthcare, sovereignty is broken into three layers, data, model and operations, and its essence is defined as the freedom of choice to swap out whatever you curre
A case of the competitive focus of embedding models moving from scores to operating costs. The design in which tiers share an embedding space aims to eliminate reindexing costs — the biggest barrier to switching models i
The first publicly disclosed large-scale extraction attempt since hiding reasoning traces became a trade-secret protection measure. Because the attribution points to a specific Chinese lab, it reads as a geopolitical eve
The first case of a competing lab putting numbers on what leaks out when open weights are released without safety measures. The debate in the open-weight lineage (axis ⑨) shifts from a "performance gap" to a "safeguard g
The key is that speed has been split off into a separately billed tier. When latency becomes the bottleneck of agent loops, tokens per second become a product separate from quality, and investment in inference infrastruc
Disaggregated serving has come down from a research claim to an operational guide. The key is using p99 latency rather than average throughput as the yardstick, since agent loops are directly exposed to tail latency. It
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
An approach that retrieves and adapts skills from repositories instead of learning them anew for each harness, turning skills from outputs tied to a model and harness into transferable assets. Read together with ClawScan
The first major commercial product to sell an agent not as "a session that runs when asked" but as "a persona resident in the account". The billing unit shifts from tokens and seats to the number of agents, and always-on
A path has opened for agent runtimes to move from vendor products to foundation-owned standard infrastructure. The key is defining the harness, model and sandbox as swappable slots: if this separation takes hold, the cos
A record of the focus of safety discussion moving from deployment time to training time. If labs accept in public documents the premise that models can escape containment during training, a procedural standard emerges th
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
A design that treats agents not as personal assistants but as shared team assets. Bundling context, credentials and memory per bot means organizational knowledge builds up in the bot, but the bot in turn becomes a concen
This bundle of reports asks whether the smart-glasses craze is a rehash of the earlier virtual-world boom, and traces in graphs how the label was swapped for spatial computing. The key causal link is put down to the mode
This shows, screen by screen, a setup that puts a text model unable to produce scenes on its own at the command line and directs it through conversation to handle everything on an external generation platform, from loggi
A commercial implementation that moves prompt-injection defense down from model alignment to the network boundary. Starting from the premise that agents cannot be kept from being fooled, it keeps data from getting out ev
The operational foundation for agents moving from "running when a person calls them" to "waking up on their own" has become a GA cloud product. It matters that the option of giving agents their own identity came out at t
This narrated briefing centers on the paradox that leading companies actually block their own employees from using outside models, and looks for the reasons in log retention periods and data sovereignty. The issue narrow
This is a beginner's walkthrough by a business owner, who assumes they can't code, that compresses everything from installing a commercial coding agent to its first task into 10 minutes. Two pieces are worth borrowing: s
This is a record of having two commercial models build games in three genres from the same one-line instruction, then playing the finished products all the way through by hand to compare them. What is worth taking away i
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
This is a write-up of a 48-minute recording that follows automation's move into the physical world through four areas: semiconductors, compute facilities, humanoids and factories. The sharpest distinction is tolerance fo
This briefly sums up the departure of a safety researcher at a frontier AI developer who left the company warning of the possibility of humanity's destruction. What deserves attention is not the departing researcher's wa
This conversation answers the conventional wisdom about the race to secure accelerators, that whoever has the most wins, by agreeing with the premise but saying that is no longer the case. The example given as evidence i
After the maker of a professional production tool officially released a command-line interface, an integration module and a skills repository holding domain knowledge, this sets out step by step the procedure for attachi
This reads several Korean news reports on robots and physical-world automation together as a single indicator. What they have in common is not the unveiling of new prototypes but that words such as sales, parts productio
This is a record of someone who never learned a professional 3D production tool using it for a day with nothing but an integration interface attached. It pins down the reason they had put it off not as a lack of skill bu
This is a side-by-side record in which the same three tasks were given to two commercial models with identical instructions, and the outputs, time taken and cost equivalent were published together for each round. In a co
All four teams' models were judged to meet the independence criteria; with a newly introduced public user evaluation included, the field narrowed to three
Brought the project back to four teams after the first-round cut, all to be evaluated in August under the same development period, GPUs and data
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
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%
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
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
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
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
Company-reported 74.5% on SWE-bench Verified; priced the same as Opus 4, with Anthropic recommending the upgrade for all uses
Unlike A.X 4.0, which continued pretraining open-source models on Korean data, it was designed and trained entirely in-house; SKT's open models now numbered four, two A.X 3.1 and two A.X 4.0
Alibaba claimed state-of-the-art results among open models in agentic coding, browser use and tool use, and open-sourced the command-line coding tool Qwen Code alongside it
The first Kanana release under a license that allows commercial use (Apache 2.0)
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
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 full release following QwQ-32B-Preview; Alibaba claims performance comparable to DeepSeek-R1 (671B total, 37B active) at 32B
The first open weights in the Kanana line, released alongside a report on training the 2.1B–32.5B Kanana lineup
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
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
Mistral recommitted to Apache 2.0 for its general-purpose models while moving away from MRL-licensed ones; company-reported over 81% on MMLU at 150 tokens/s
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
A follow-up four months after the EXAONE 3.0 7.8B release that expanded to three sizes while keeping a non-commercial license
An update to the Llama 3 line that shipped a single 70B instruction-tuned model, with no separate pretrained release
The 11B and 90B were the first Llama models to support image input, released alongside 1B and 3B models aimed at on-device use
All open models except the 3B and 72B were released under Apache 2.0, and the 14B and 32B sizes missing from the previous generation returned
A preview released as weights plus free API access ahead of the full release planned for November; Upstage claimed it competes with multi-GPU models such as Llama 3.1 70B
Published the weights of a 123B flagship designed for single-node inference, limited to research and non-commercial use, with a commercial license required for self-deployment
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
The second Gemma generation; Google claimed the 27B competes with models more than twice its size and the 9B outperforms Llama 3 8B
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
An open-weight code model that DeepSeek claimed matched GPT-4 Turbo on code-specific tasks
Only the 72B kept the original Qianwen License while the other four moved to Apache 2.0, the release where Qwen's licensing began to open up
Introduced a non-production license that publishes weights for research and testing only, with commercial use requiring a separate license
Mistral's larger MoE follow-up to Mixtral 8x7B, again under Apache 2.0
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 API flagship launched with Microsoft Azure as Mistral's first distribution partner; Mistral claimed it ranked second only to GPT-4 among models available through an API
A model from a Korean startup that, per Upstage, topped the Hugging Face Open LLM Leaderboard at release; the 74.2 average belonged to the instruction-tuned SOLAR-10.7B-Instruct, while the base model scored 66.04
A general-purpose DeepSeek model released down to its intermediate checkpoints; DeepSeek claimed the 67B Base outperformed Llama 2 70B Base in reasoning, coding, math and Chinese comprehension
A code-specialized line trained on 500B tokens of code and code-related data on top of Llama 2; a 70B version was added on January 29, 2024
Company-reported scores: 58.7% on Codex (Instant 1.1: 52.8%) and 86.7% on GSM8K (80.9%); Anthropic said it hallucinates less and is more resistant to jailbreaks
Alibaba's first open-weight language model; the original release was trained on 2.2T tokens, raised to 2.4T in an update on September 25, 2023
Company-reported scores: Codex HumanEval 71.2% (Claude 1.3: 56.0%), GSM8k 88.0% (85.2%), and 76.5% on the multiple-choice section of the Bar exam (73.0%); API priced the same as Claude 1.3
At Google I/O, Google announced over 25 products and features powered by PaLM 2, including Bard, and said the smallest size, Gecko, can run on mobile devices even offline
Google said it would start with a lightweight version of LaMDA to reduce compute needs and gather feedback, before wider public availability in the coming weeks
The first large-scale use of the Pathways system; reported surpassing the few-shot performance of prior large models such as GPT-3, Gopher and Chinchilla on 28 of 29 English NLP tasks
Presented by Naver as Korea's first hyperscale AI and a means to AI sovereignty; a closed model later succeeded by HyperCLOVA X
Unlike most language models it was trained on dialogue, and Google said fine-tuning after training could significantly improve the sensibleness and specificity of its responses
Fine-tuned with just one additional output layer, it set new state-of-the-art results on 11 NLP tasks, including GLUE 80.4% and SQuAD v1.1 F1 93.2 (as reported by the authors)
Reported that combining unsupervised generative pre-training with supervised fine-tuning improved the state of the art on 9 of the 12 tasks studied