LLMs

Latest stories covering LLMs, with summaries and links to related coverage.

22 stories · 22 reports · page 1 of 1

Latest stories

• There are now 5 inference providers for abliterated DeepSeek

  2026-10-27

  tool

  1 min read

  AllModels routes abliterated language models across providers, offering automatic fallbacks when a selected provider is unavailable. The service uses modality-aware selection to match routing with model capabilities. AllModels also provides a single place to compare model availability and token prices across providers.

  Model routingAI agentsDeveloper tools

• GPT-6.1 Terra

  2026-10-26

  tool

  4 min read

  OpenAI has introduced GPT-6.1 Terra, a model for visual reasoning, computer-use tasks, writing, and coding. It is available immediately to ChatGPT Work and Codex users on Plus, Pro, Enterprise, and Edu plans, and through the OpenAI API as gpt-6.1-terra. OpenAI lists standard API prices of $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens; it says these prices are 50% below GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says GPT-6.1 Terra can handle tasks including interpreting technical drawings, entering invoice data into accounting software, completing forms, and editing video. On the Agents’ Last Exam computer-use benchmark at medium reasoning effort, OpenAI reports that Terra scored 15 percentage points above GPT-5.6 Terra at about one-third the cost, while matching GPT-6 Astra’s performance at one-tenth its token price. It says Terra completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. OpenAI reports a lower error rate on its internal error-inducing-prompt benchmark: 16.7% at low reasoning effort, versus 53.6% for GPT-5.6 Terra, and 4.3% at maximum effort. On Terminal-Bench 4.0, Terra solved 50.3% of tasks at maximum reasoning effort, compared with GPT-5.6 Terra’s published 21.5%. OpenAI says the model meets its Critical cybersecurity threshold and retains GPT-6 Astra’s safeguards.

  GPT 6OpenAIAI agents

• Sebastian Raschka's from-scratch video course on building reasoning models

  2026-10-26

  opinion

  11 min read

  An eight-video, 11-hour coding course teaches developers to build and improve a small reasoning model in Python and PyTorch, beginning with a pre-trained Qwen3 0.6B model. The materials cover model loading, streaming text generation, KV caching, model compilation, evaluation, inference-time scaling, reinforcement learning, and distillation. Python and PyTorch familiarity are recommended; early generation and evaluation exercises can run on a CPU, while later training experiments are more practical on a GPU. The course builds a math-answer verifier and uses the MATH-500 benchmark to evaluate models. It implements temperature scaling, multinomial sampling, top-p filtering, self-consistency voting, scoring based on token and sequence log probabilities, and self-refinement. Later lessons implement a simplified Group Relative Policy Optimization method for reinforcement learning with verifiable rewards, including response groups scored by the verifier, training metrics, clipped policy ratios, and KL-loss terms. The final lesson distills reasoning behavior from DeepSeek R1-generated answers through supervised training and compares distilled-model results on MATH-500. The videos accompany the book Build a Reasoning Model (From Scratch), and a separate 90-minute LinkedIn Learning course provides a conceptual overview of inference-time scaling, reinforcement learning, and distillation.

  Reasoning modelsReinforcement learningAI education

• Anthropic S-1

  2026-10-26

  news

  1 min read

  An SEC filing link identifies Anthropic, PBC and an S-1 document labeled “Q2.” The source provides no filing details, financial figures, dates, business information, or statements from Anthropic. The link was posted by SiNTEx and had a score of 1,064 points and 989 comments at the time reflected in the source.

  AnthropicAI agentsDeveloper tools

• OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?

  2026-10-25

  news

  11 min read

  OpenAI and Anthropic cut prices for several leading language models between June and September 2026, with reductions reaching 80% for some services. OpenAI lowered GPT-5.6 Luna from $1/$6/$0.10 to $0.20/$1.20/$0.02 per million input, output, and cache-read tokens on July 30, then launched GPT-6 Luna at $0.10/$0.50/$0.01 on September 22. GPT-6.1 Sol, released September 29 after GPT-6 Sol drew weak performance feedback, cost $2/$10/$0.10 and was described as comparable in performance to GPT-6 Astra. Anthropic made Claude Sonnet 5’s $2/$10/$0.20 launch pricing permanent and priced Claude Opus 5.5 at $4/$20/$0.20, saying the model requires less compute to serve than Opus 5. The source attributes the cuts partly to direct competition and pressure from open-weight models and numerous inference hosts. DeepSeek Flash V4.1 was priced by DeepSeek at $0.15/$0.60/$0.003, while OpenRouter listed 29 providers for the model’s updated variant. The source says cache-read charges dominate token spending in coding-agent workloads, estimating that cache reads account for 96% to 99% of expenditure; Anthropic’s 75% cache-read reduction for Fable 5.1 could therefore reduce total usage costs by roughly 70% to 75%. It also cites improved inference efficiency as a possible contributor to lower prices.

  OpenAIAnthropicAI pricing

• A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

  2026-10-25

  opinion

  5 min read

  A dermatologist tested GPT-6 Astra, Claude Fable 5.1 and a fine-tuned Inception v3 convolutional neural network on 200 randomly selected, biopsy-confirmed skin lesions from the public MILK10k dataset. The CNN correctly identified 125 of 200 lesions across six diagnoses, or 62%, while GPT-6 Astra and Claude Fable 5.1 identified 84 and 85, or 42% each. Inception v3 was introduced by Google in 2015 and was fine-tuned on 68,011 ISIC Archive clinical and dermoscopic images for about 20 minutes on one H100 GPU. The test set contained 100 lesions requiring further care, including 30 melanomas, and 100 benign lesions considered safe to discharge. The CNN identified 13 melanomas as its first choice, versus eight for GPT-6 Astra and 11 for Claude Fable 5.1. With a second diagnostic choice allowed, the CNN reached 173 correct cases, compared with 134 and 138 for the LLMs. Both LLMs incorrectly labeled a pictured melanoma as a benign seborrhoeic keratosis and said the patient could be discharged. The sample size was small, each LLM response was generated once, and all lesions had been biopsied, making the cases unusually difficult. At its preset discharge threshold, the CNN discharged 18 melanomas, though matching its unsafe-lesion retention rate to the LLMs reduced that figure to two or one. The browser-based CNN runs locally with WebGPU and does not upload images.

  Computer visionHealthcare AIAI agents

• AI use leads UC Berkeley math professor to see 'no point' in grading homework

  2026-10-25

  news

  2 min read

  UC Berkeley mathematics professor and Fields Medalist Richard Borcherds has stopped grading homework for correctness after seeing students use AI to consistently earn top scores. He now gives credit for completion when students submit assignments, calling the arrangement “not a very good system” but saying he has no alternative while AI use is unavoidable. Borcherds said there was no point in paying readers to assess homework correctness under those conditions. In an interview following OpenAI’s announcement that AI agents had apparently solved the Navier-Stokes existence and smoothness problem, Borcherds said that, if AI improves at its current rate, it could soon outperform human mathematicians in every task. He said he no longer knows whether to encourage students to enter mathematics and that even ChatGPT did not offer him a clear prediction about the profession’s future. He expects academic inertia could leave math departments largely unchanged for several years despite AI advances. UC Berkeley asks faculty to include course expectations for AI use in syllabi, but individual instructors set class policies. The campus has tested AI-detection tools including Turnitin and Rumi, but has concluded they are not reliable enough for campuswide standardization. Berkeley’s AI Hub provides faculty with sample syllabus statements and guidance on AI tools and policies.

  AI agentsOpenAIEducation technology

• Cerebras Deploys Trillion-Parameter AI Models

  2026-10-25

  research

  6 min read

  Cerebras says it can deploy multi-trillion-parameter AI models by distributing their layers across WSE-3T wafer-scale chips and connecting the chips as an inference pipeline. Each WSE-3T wafer contains 44 GB of on-chip SRAM, while a trillion-parameter model requires terabytes of weight storage. Cerebras says this layout can run multi-trillion-parameter models at 1,000 to more than 2,000 tokens per second. Using the open-source Kimi K3 model as an example, Cerebras maps its roughly 3 trillion parameters, 69 linear-attention layers, and 24 dense layers across 59 wafers. With 4-bit weight quantization, each layer occupies about 16 GB. Dense layers are assigned one per wafer because the remaining memory is used for the key-value cache; two linear-attention layers fit on a wafer because they use fixed-size per-request state rather than a context-length-dependent KV cache. Activations, rather than model weights, move between wafers through RDMA over Ethernet at roughly two microseconds per hop. Multiple requests remain in flight simultaneously through pipeline stages, and each stage can process a micro-batch. Cerebras says smaller micro-batches reduce individual-request latency, while larger ones raise throughput. Available KV-cache memory on dense-layer wafers limits concurrency for 256,000-token average requests using an 8-bit KV cache.

  Trillion parameter modelsCerebrasAI hardware

• Why is AI generated language so hard to understand?

  2026-10-24

  research

  2 min read

  A study submitted to arXiv on October 21, 2026 finds that AI-generated text uses several linguistic features associated with greater human processing difficulty more often than comparison texts. The researchers analyzed a new corpus of AI writing from diverse domains and measured features including object relative clauses, long syntactic dependencies, and low-frequency word sequences. AI models used difficult features more often than paraphrases produced through round-trip translation. Against human writing, all nine frontier and near-frontier models examined used more inanimate subjects paired with animate verbs and more low-frequency n-grams. Most models also produced longer dependencies, more stacked nouns, and more reduced object relatives in conversational text, although the results varied substantially among model families. The researchers propose that human language is shaped for successful communication, while AI output is increasingly shaped by objectives such as mathematics, code, and reasoning that reward information density. As a proof of concept, fine-tuning a language model with a length penalty caused its language behavior to diverge from human-like patterns. The findings identify potential mismatches between the linguistic demands of human readers and language-model output and suggest that model training objectives may contribute to harder-to-understand text.

  AI generated textPsycholinguisticsLanguage processing

• New AI Is More Than Twice as Good at Diagnosing Diseases as Doctors

  2026-10-23

  research

  3 min read

  Microsoft’s AI Diagnostic Orchestrator, or MAI-DxO, achieved 83.2% diagnostic accuracy on complex cases from The New England Journal of Medicine, compared with 36.1% for generalist physicians using reference materials. Physicians without those materials scored about 22%. In a separate emergency-department admissions test, MAI-DxO scored 80.3%, versus 46.8% for generalist physicians. Microsoft reported its strongest results with OpenAI’s GPT-5 and additional improvements with GPT-5.5. MAI-DxO begins with a patient’s chief complaint, then sequentially selects follow-up questions and diagnostic tests before reaching a conclusion. It assigns AI agents roles including proposing diagnoses, choosing tests, challenging assumptions, and weighing costs. Microsoft said simulated-case estimates indicated the system reached diagnoses using fewer tests and procedures at roughly half to one-third of the cost, though those estimates were not actual medical expenses. The findings, first posted as an arXiv preprint last year, have been published in Science. Microsoft AI chief Mustafa Suleyman said real-world use would require clinical testing. Dr. Dominic King, Microsoft’s AI health lead, said testing was expected to begin at Beth Israel Deaconess Medical Center in Boston within months. The researchers said the performance gap would likely narrow against specialist physicians rather than generalists, and medical experts have said diagnostic benchmark performance does not establish that AI should replace doctors.

  Healthcare AIMicrosoftAI diagnostics

• Gemma-4-26B-A4B with 39-69% less thinking at same accuracy

  2026-10-23

  research

  8 min read

  Bottlecap AI released ThinkingCap-gemma-4-26B-A4B-it, a token-efficient version of Google’s Gemma-4-26B-A4B-it reasoning model. Gemma-4-26B-A4B is a sparse mixture-of-experts model with 26 billion parameters and about 4 billion active parameters per token. Bottlecap says its model uses 38.9% fewer thinking tokens on average across 12 benchmarks while accuracy declined by 0.15 percentage points, from 75.18% to 75.03%, under identical vLLM serving and sampling settings. Thinking-token reductions ranged from 13.7% on IFBench to 69.1% on LongBench v2. The model used 54.4% fewer tokens across long-context and multimodal tasks, 44.4% fewer on knowledge and reasoning, 36.1% fewer on math, and 27.6% fewer on code. Accuracy increased on some tests, including IFBench and LongBench v2, but fell 4.00 percentage points on AA-LCR. Bottlecap measured looping in 0.76% of the optimized model’s responses, down from 3.64% for the base model. Bottlecap evaluated both models with thinking enabled, temperature 1.0, top-p 0.95, top-k 64, and a 65,536-token generation cap. The open model and GGUF builds are available under the bottlecapai/ThinkingCap-gemma-4-26B-A4B-it repository. Bottlecap also offers a commercial PRO version, which it says cuts thinking tokens by up to 87.6% and 62% on average.

  ThinkingcapGoogleAI efficiency

• LLMs are making formal software verification feasible at a much larger scale

  2026-10-22

  research

  17 min read

  Anthropic and formal-verification company Theorem used Claude to generate machine-checkable proofs for security properties in open-source software at the RISC-V assembly level. Over several weeks, Claude patched bugs and proved that 239 OSS-Fuzz entry points across 123 libraries would not crash for any input below 64 MB, under assumptions including a single-threaded static RISC-V binary, Linux, and a fixed memory budget. The proofs cover normal exits without assertion failures or modeled memory-safety errors, rather than proving that the libraries are entirely bug-free. The effort produced roughly 180 million lines of Rocq proof code, with a typical proof about 600,000 lines for a binary containing a median 18,000 RISC-V instructions. Claude found hundreds of bugs while attempting the proofs, including out-of-bounds reads and writes requiring unusual or large inputs, a rediscovery of CVE-2026-56132 in libexpat, and a logic flaw in a file-loading restriction feature. Anthropic is publishing a subset of the proofs, relevant RISC-V and Linux semantics, and specifications that it could not prove; Theorem plans to offer the pipeline for other open-source projects. The guarantees depend on the correctness and scope of the formal specification, processor and operating-system models, and the Rocq proof checker. The current proofs do not model common deployments such as x86 or ARM systems with dynamically linked glibc, and AddressSanitizer-based fuzz targets do not capture every possible vulnerability.

  Formal verificationAI safetySoftware development

• How does Fable 5.1 hide messages in images?

  2026-10-22

  research

  17 min read

  Tests of Claude Fable 5.1 found that the model can hide short messages in ordinary-looking SVG artwork and that a separate Fable session can often recover them from a rendered PNG without seeing the SVG source code or receiving a prearranged cipher. Across 50 messages in 15 languages, Fable most often assigned alphabet positions to the heights of repeated visual elements: 33 images used height encoding, frequently as buildings in night skylines. Other images encoded data through star positions, grass-blade heights, Morse-like dashed trails, and lit or unlit windows. A separate Fable 5.1 instance, run at maximum reasoning effort with Python, Pillow, NumPy, and SciPy tools, recovered 78% of messages under an automated 90%-similarity threshold. It inspected pixels and geometric patterns, such as star coordinates that mapped to letters or blade heights that represented alphabet indices. Fable also embedded contextual hints for non-English alphabets, accents, capitalization, and digits. GPT-6 Astra used different schemes in the same 50-message test, favoring Morse code in short and long objects in 18 cases, plus acrostics, bit codes, and Braille-like window panes. In cross-model tests, Astra recovered 44 of Fable’s 50 messages, while Fable recovered 42 of Astra’s. The researchers say decoding currently leaves visible evidence in tool calls and do not claim the technique is presently dangerous.

  ClaudeImage generationAI safety

• Every AI model has the same "controversial" take on humans

  2026-10-22

  opinion

  3 min read

  ModelCompare prompted 12 text AI models three times each to state a controversial opinion about humans. In 29 of 35 recorded responses, the models characterized people as prioritizing group belonging, status or loyalty over truth and principles. Seven of the 12 models gave that view in all three responses. The tested systems came from 10 labs in three countries and included GPT-6.1 Sol, Gemini 4 Argon, DeepSeek V4 Pro, Grok 4.7, Claude Opus 5.5 and Mistral Large 4. One Grok response declined to provide a controversial opinion. Several responses differed from the dominant theme. Claude Opus 5.5 said people are more decent than they believe and cited lost wallets being returned more often than expected. Claude Fable 5.1 focused on intentions behind harm and help, GLM 5.3 said people may sacrifice comfort to feel significant, and Mistral Large 4 called consciousness an evolutionary malfunction. Six image models also received two prompts each to create satirical caricatures of humans. Ten of 12 images portrayed humans as phone-dependent consumers damaging the planet, including scenes of people scrolling off cliffs or taking selfies during disaster. None of the text-model responses mentioned phones or environmental damage. ModelCompare said all 60 text and image calls cost $1.54; text responses were manually grouped, and Nano Banana 2.1 rendered the text models’ cartoon briefs.

  AI modelsHuman behaviorAI opinions

• Mistral Large 4

  2026-10-21

  tool

  11 min read

  Mistral has launched a public-preview API for Mistral Large 4, a natively multimodal mixture-of-experts model with 1 trillion total parameters and 52 billion active parameters. The company plans to release its open weights by the end of the month. Mistral Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral-operated European data centers, using multilingual data spanning more than 160 languages. The preview API is available through Mistral Studio, priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the model is its largest and most capable release, targeting coding, tool-using agents, cybersecurity, finance, law, science, and visual understanding. It reported scores of 82% on a vulnerability reproduction-and-patching test, 93% on Cybench security challenges, 61.7% on DeepSWE v1.1, 59.9% on AutomationBench, and 42% on Dense 200 visual grounding. Mistral also said ML4 resists 93.3% of attacks in Lakera’s B3 AI Security Benchmark. Before the weight release, Mistral is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities that will receive reduced moderation and expanded cyber capabilities. The company says ML4 can be self-deployed in private clouds or on premises and will underpin specialized future Mistral models. Its development is funded in part by Mistral’s €3 billion Series D round.

  MistralMultimodal modelsAI performance

• On AI writing

  2026-10-20

  opinion

  2 min read

  AI-generated writing can obscure an author’s distinctive style, reduce precision, and shift time costs from writers to readers, according to the author. The author’s bachelor’s thesis found that word-use statistics from short passages could identify authors reasonably well, while readers also recognize patterns in sentence structure and storytelling. The author argues that AI tends to average stylistic traits, making it harder for readers to infer a writer’s personality or context and to empathize with them. The author says language models must make assumptions to connect ideas, even when prompts are detailed, and may therefore choose broader or more ambiguous wording to avoid logical errors. Writers may miss that ambiguity because they supply intended meanings from their own understanding. The author also argues that drafting independently helps expose gaps and inconsistencies in a person’s thinking, whereas AI can produce a finished-seeming result before that scrutiny occurs. Ambiguous language and reduced stylistic context can make AI-generated text slower to read, the author says. The author estimates reading AI text takes at least one-third longer for them than text of the same length written without AI, and says many models’ verbosity further increases reading time. Any writing-time savings can disappear when multiple readers spend longer interpreting the output, particularly when ambiguity produces follow-up questions, revisions, and lower trust.

  AI writingText analysisAuthor attribution

• Publish fake stats, get cited in LLMs

  2026-10-20

  opinion

  13 min read

  StateGlobe, a website launched in March 2026, published 50,000 AI-generated and unverified statistics covering SEO, marketing, e-commerce and web technology across 200 countries. The site, created with GPT-4.1 for roughly $50 and later expanded with 15,000 listicles, was largely deindexed by Google after launch but continued to receive retrieval traffic from ChatGPT. Over seven months, GA4 recorded more than 10,000 sessions from ChatGPT, while Cloudflare logged about 72,000 ChatGPT-User requests in the most recent 30 days and more than 600,000 bot requests overall. According to Ahrefs and Semrush, more than 330 referring domains linked to StateGlobe, including Shopify, DHL and Bluehost. The site’s invented figures were cited in an academic paper, investment research from Qatar’s largest financial group, an Indian economics publication and a Canadian influencer-community report. The experiment’s operator says these citations can make false claims appear more credible as other writers and models encounter sources that cite them. Peec AI subsequently launched a statistics page using checked data and source citations, reporting that it also began earning links quickly. The company recommends publishing verifiable, passage-level statistics with adjacent sources and dates, query-aligned headings, visible update dates, and accessible server-rendered HTML. It warns against publishing fabricated material and says LLM retrieval systems may change.

  AI searchChatGPTData ethics

• Inception "Mercury 3", 1000 tok/s diffusion LLM

  2026-10-19

  tool

  4 min read

  Inception Labs has launched Mercury 3, a diffusion-based large language model that it says delivers more than 1,000 tokens per second on Nvidia GPUs while matching or exceeding frontier speed-optimized models on benchmarks. The company reports a measured speed of 1,068 tokens per second, a one-million-token context window, and a model more than 40 times larger than Mercury 1. Inception says Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks and that its benchmark intelligence rose threefold from Mercury 1 mini. Mercury 3 is available through Inception’s API platform, OpenRouter, Baseten, and Vercel, with OpenAI API compatibility. Pricing is $0.20 per million input tokens and $0.70 per million output tokens, reduced at launch to $0.10 and $0.35 respectively; cached input tokens receive a 90% discount from the standard input price. Enterprise options include dedicated capacity, autoscaling, compliance controls, and configurable data retention. Early-access users deploy the model for browser agents, tutoring, search, and customer support. Inception says Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in its web-agent harness, exceeding GPT-6.1-Sol under the same token volume and agent-step count. Octen AI reported median generation time of 0.9 seconds and median end-to-end response time of 2.2 seconds in Sense API tests.

  Mercury 3AI agentsText diffusion

• Harness Engineering 101: The LLM is the brain, the harness is the body

  2026-10-19

  tool

  4 min read

  Harness Engineering 101 is a September 2026 book and GitHub repository by Isuru Wijesiri that explains how to build AI agents from first principles. It defines a large language model as the “brain” and the harness as the programmatic “body” that builds, maintains, and protects the conversation history sent to a stateless LLM API. Each API turn sends the full conversation as a JSON array and receives generated text in return. The book develops a roughly 300-line Python toy agent across 16 chapters, with each chapter presenting a failure, a minimal patch, and the patch’s cost. Topics include tool calling, repeated agent loops, caching, context-window limits, subagents, steering, asynchronous background work, debugging, guardrails, coding agents, retrieval-augmented generation, MCP, skills, hooks, model routing, worktree isolation, operating modes, retries, rate limits, and streaming. Wijesiri says the same ideas are used in One Code, an open-source coding agent built on pi. The repository includes Markdown book sources, runnable toy-harness stages, diagram generators, and mdBook configuration. The prose is licensed under CC BY-NC-ND 4.0, while Python harness and diagram code is MIT-licensed. The book can be built locally with mdBook and viewed through a live-reloading server.

  AI agentsDeveloper toolsHarness engineering

• DeepSeek V4.1 Flash 1017 (Max)

  2026-10-19

  research

  5 min read

  DeepSeek released the reasoning model DeepSeek V4.1 Flash 1017 (Max) on October 17, 2026. The open-weights model accepts text and image inputs, produces text, and supports a 1 million-token context window. It is a mixture-of-experts model with 552 billion total parameters and 16 billion active parameters during inference. Its weights are available on Hugging Face under the MIT license, which permits commercial use, and it is available through five API providers. Artificial Analysis gave the model a score of 45 on its Intelligence Index, compared with a median score of 19 among open-weight models of a similar size. The composite index evaluates reasoning, knowledge, mathematics and coding. The evaluation generated 210 million output tokens, above the comparable-model median of 140 million, indicating relatively high verbosity. DeepSeek’s API pricing is $0.30 per million input tokens and $1.20 per million output tokens, below respective comparison medians of $0.43 and $1.68. Artificial Analysis estimates an average Intelligence Index evaluation cost of $0.26 per task. Output speed measured 230.4 tokens per second, compared with a 77.9-token-per-second median, while time to first token was 1.08 seconds versus a 1.69-second median.

  AI performanceModel evaluationText and image processing

• If you don't run your own LLM proxy, how do you know whats going on?

  2026-10-18

  opinion

  2 min read

  Running a self-hosted LLM proxy can centralize requests to multiple AI providers and models behind an endpoint the operator controls. The setup can log which API key called which model, which client made the request, the request time, and its status. Virtual API keys can limit the spread of a primary provider key across applications, agents, and test harnesses, while a proxy can allow rapid switches between models or providers without reconfiguring every client. The writer argues that consumer-facing AI usage dashboards do not provide the time-series token data needed for engineering decisions. They use a modified self-hosted proxy to add visibility, custom Model Context Protocol server endpoints, and other features. The proxy’s API could also support automated diagnostics; the writer says they can manually identify the cause of a cache miss during a coding session in about 20 seconds. They run this infrastructure at home for learning and at work to maintain control over application visibility, data paths, storage, and inference instrumentation.

  Developer toolsAI agentsAutomation

• Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

  2026-10-18

  research

  2 min read

  Hybrid Reasoning Agents, submitted to arXiv on October 13, 2026, combine latent reasoning with text-based interaction in a single language-model agent. The system is designed to let an agent use latent representations for internal reasoning while continuing to use text to call tools and interact with external environments. Its latent thoughts are intended to remain directly inspectable. The architecture uses a lightweight component called HybridLink to switch between text and latent modes during generation. A continuous working-memory mechanism retains hybrid context across multiple turns. The researchers also describe two post-training methods: HybridSFT, which trains a backbone model to coordinate with HybridLink during tool-integrated generation, and HybridRL, which learns when to reason in latent space and when to act through text. Across three model scales and 11 benchmarks in mathematics, science, medicine, search, and agentic reasoning, the authors report that HybridAgent improved performance by 10.4% relative to frontier LLM agents and agentic frameworks using text-based or latent methods. They also report 35.5% to 54.7% lower token use, 1.8× to 2.6× faster inference, and 2.7× to 3.7× cost savings. The authors state that the project’s code and data are available online.

  Hybrid reasoningAI agentsDeveloper tools