Coverage of the large language model frontier: frontier model releases, benchmarks and evaluations, research papers worth reading, and the infrastructure work behind million-token context windows.
Topic brief
Brief updated Aug 3, 2026
Large language models are neural networks trained on vast text and multimodal corpora to predict and generate language, and they are the core engine of modern generative AI. Foundation and frontier models from a handful of labs now power chat assistants, coding tools, agents, search and enterprise automation, which makes their capabilities, costs and access terms a central concern for anyone building with AI.
For practitioners, LLMs are both a capability and a moving target. Each frontier release shifts benchmarks, token pricing, context length and tool-use reliability, and those shifts ripple through product design, infrastructure budgets and vendor strategy. The economics matter as much as the intelligence: inference cost per token, distillation, quantization behavior and the split between commodity and frontier tiers determine what is affordable to deploy at scale. Model behavior under evaluation is also becoming a first-class concern, since models that optimize against the scoring mechanism rather than the task distort the numbers teams rely on.
Competition here is geopolitical and increasingly regulated. Model weights are strategic assets subject to export controls, open-weight releases from Chinese and other labs are reshaping pricing and developer share, and governments now test frontier systems for safety and cyber risk before broad release. Where a model runs is a third axis: the same weights can be served from rented GPUs, from custom accelerators, or compressed onto a laptop or a phone, and the choice changes cost, latency and compliance posture. Tracking LLMs means tracking not just which model is smartest, but which is available, affordable, compliant and safe to run.
The through-line across late July and early August 2026 is that top-end capability, self-hostable scale and verifiable evidence are now moving on three separate tracks. OpenAI disclosed on August 1 that an internal version of Astra, which it describes as its next major model, produced ten results in mathematics and theoretical computer science, and it published human-prepared manuscripts, a Lean formalization of each argument and a model narration of the reasoning, estimating the total token cost of finding the ten solutions at roughly $2,000 at Sol API rates. That release shipped inspectable proof artifacts instead of benchmark scores, but OpenAI published no model card, architecture, context-window specification, API details or release date for Astra. In the same three-week window the open-weight tier reached trillion-parameter scale: Moonshot AI published Kimi K3's full weights on July 27, with a model card listing 2.8 trillion total parameters, 104 billion activated, 896 experts with 16 selected per token and a 1,048,576-token context window under the Kimi K3 License, and Simon Willison put the Hugging Face repository at roughly 1.56 TB. Thinking Machines Lab released Inkling on July 15 at 975 billion total and 41 billion active parameters under Apache 2.0, and LG AI Research published K-EXAONE 2.0 on July 31 at 750 billion total and 37 billion active parameters, also Apache 2.0. Weights at that size shift infrastructure, evaluation, security review and monitoring costs onto whoever downloads them; Nature reported that K3's size alone could limit adoption even if its capabilities hold up.
Pricing and serving economics moved in the same direction at the same time. Anthropic released Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8's base pricing and half the price of Claude Fable 5, and made it the default for Claude Max. OpenAI then cut GPT-5.6 Terra by 20% and Luna by 80% on July 30, listing Terra at $2 and $12 and Luna at $0.20 and $1.20 per million input and output tokens, and said on July 29 that Sol-assisted production-kernel work had reduced its own end-to-end serving costs by 20% while a Sol-improved speculative-decoding draft model raised token-generation efficiency by more than 15%, both company-reported production measurements with no announced customer price change attached. Cheaper tokens and downloadable weights do not reduce the verification burden, they concentrate it. DeepSeek's July 31 V4-Flash-0731 release is the clearest example: its own model card reports 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE using an unreleased minimal mode of DeepSeek Harness and internal test sets, while Artificial Analysis independently scored the model 50 on its Intelligence Index, up from 40, and found its hallucination rate falling 11 points to 84% with accuracy unchanged at 37%. Arena, whose factuality leaderboard rests on more than 2 million labeled claims from real conversations weighted 25% factuality against 75% human preference, told Lets Data Science that OpenAI is the only provider consistently improving factuality over an extended period and that most open-source models lose ground once factual accuracy is counted. For anyone building or buying, the operative response is unchanged and more urgent: pin model versions, reproduce vendor benchmarks inside your own harness, and price workloads on measured utilization rather than list rates.
Several commitments in this evidence are dated and still open. DeepSeek says Responses API support for V4 Pro is expected in early August 2026, and its published code-agent scores rest on an unreleased minimal mode of DeepSeek Harness plus internal DSBench-FullStack and DSBench-Hard test sets, so watch whether that evaluation infrastructure is released. OpenAI has published no model card, architecture, context-window specification, API details or release date for Astra, and Gizmodo reported on August 2 that the company had not replied to questions about the model's official name; watch whether independent mathematicians inspect the Lean formalizations. Google says Gemini 4 has entered its most ambitious pretraining run but has disclosed no release date, parameter count, context window, pricing or reproducible coding evaluations. Reuters reported that MiniMax is developing a 2.7-trillion-parameter open-weight model that could arrive as early as the third quarter of 2026, with the company declining to comment, and Sarvam AI said on July 30 that it plans a trillion-parameter model, with the Free Press Journal reporting a six-month expectation and no architecture, training-compute plan or evaluation methodology published. On policy, the evidence does not establish that any binding regime exists: Bloomberg reports the proposed FINRA-like AI regulator reporting to the SEC remains unpublished and had not been reviewed by President Trump, Axios reported that Commerce expects to announce a permanent CAISI director in the coming weeks following Chris Fall's resignation, and Rest of World reported that China's Commerce Ministry had convened major Chinese technology companies to discuss controls on releases of their most advanced models. Finally, watch whether any third party reproduces Moonshot's Kimi K3 benchmark table, which the retrieved coverage notes mixes company-run tests, public leaderboard results and different agent harnesses.
| model | license | total parameters | active parameters |
|---|---|---|---|
| Kimi K3 (Moonshot AI) | Kimi K3 License | 2.8 trillion | 104 billion activated |
| Inkling (Thinking Machines Lab) | Apache 2.0 | 975 billion | 41 billion per token |
| K-EXAONE 2.0 (LG AI Research) | Apache 2.0 | 750 billion | 37 billion |
| Mistral Large 3 | Apache 2.0 | 675 billion | 41 billion |
Three arrived within about two weeks. Moonshot AI's Kimi K3 model card lists 2.8 trillion total parameters, 104 billion activated, 896 experts with 16 selected per token and a 1,048,576-token context window under the Kimi K3 License; Simon Willison put the Hugging Face repository at roughly 1.56 TB, and Moonshot recommends supernode configurations with at least 64 accelerators for serving. Thinking Machines Lab's Inkling (975 billion total, 41 billion active, Apache 2.0) and LG AI Research's K-EXAONE 2.0 (750 billion total, 37 billion active, Apache 2.0) are smaller but still sparse mixture-of-experts systems whose memory footprint, expert routing and communication overhead drive cost more than the headline parameter count. Downloadable weights do not remove the operational costs of inference infrastructure, model evaluation, security review, fine-tuning and ongoing monitoring; they move them in-house.
Anthropic listed Claude Opus 5 at $5 per million input tokens and $25 per million output tokens on July 24, matching Opus 4.8's base pricing and half the price of Claude Fable 5. OpenAI cut GPT-5.6 Terra by 20% and Luna by 80% on July 30, listing Terra at $2 per million input and $12 per million output tokens and Luna at $0.20 and $1.20 respectively. Moonshot's Kimi K3 launch post lists $0.30 per million cached input tokens, $3 per million uncached input tokens and $15 per million output tokens, and SpaceXAI's docs list Grok 4.5 at $2 per 1M input tokens and $6 per 1M output tokens. These are vendor list rates; a DigitalOcean tutorial published July 8 makes the complementary point that dedicated-GPU economics depend on sustained utilization rather than hourly price alone.
Most of it, and the release notes generally say so. DeepSeek's model card reports 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE and 70.3 on Toolathlon-Verified, while stating that its public code-agent benchmarks used an unreleased minimal mode of DeepSeek Harness at max reasoning effort and that DSBench-FullStack and DSBench-Hard are internal test sets. Anthropic's Frontier-Bench v0.1 comparison for Opus 5 used the mini-SWE-agent harness, a Google Kubernetes Engine backend and five attempts per task. Moonshot's Kimi K3 table mixes company-run evaluations with cited third-party leaderboard results and different agent harnesses, and the retrieved coverage does not independently reproduce it. Independent measurement exists but is narrower: Artificial Analysis scored DeepSeek-V4-Flash-0731 at 50 on its Intelligence Index, up from 40 for the previous V4 Flash.
It depends on which failure mode matters, and the evidence separates capability from accuracy. Artificial Analysis's AA-Omniscience evaluation of DeepSeek-V4-Flash-0731 found the hallucination rate falling 11 points to 84% while accuracy stayed at 37%, a result specific to that evaluator's test design. Arena told Lets Data Science that many preferred models decline once factuality is included, with a general trend among most open-source models. On the research side, a preprint from a team affiliated with IIIT Hyderabad, TU Berlin, Microsoft Research and IIT Delhi reported that roughly 3-billion-parameter models matched 7B-14B models at predicting fMRI responses while 1B-1.5B models degraded, but also found that compression reduced performance on some discourse, syntax and morphology tasks even when the brain-prediction metric stayed stable.
A July 8, 2026 arXiv paper from Tel Aviv University, Technion and Intuit describes HalluSquatting, in which attackers pre-register repository or skill names that models predictably hallucinate; the authors report hallucinated resource generation as high as 85% in repository-cloning scenarios and 100% in skill-installation scenarios, with transfer across models, prompts and application layers. The researchers say affected vendors, model providers and marketplace maintainers were notified before publication and reproducible implementation details were redacted, so this is disclosure research rather than a confirmed exploitation campaign. Separately, Vectoral's June 28 investigation maps a relay market that resells pooled access to U.S. models, with its price comparison placing the deepest observed discount at 97.8% below official list pricing, which makes hard spend caps, concurrency limits and anomaly detection relevant controls for any model-backed public endpoint.
No binding regime is established in this evidence. Bloomberg reports the Trump administration is considering an independent AI regulator that would report to the Securities and Exchange Commission and be modeled on FINRA, developed with Treasury Secretary Scott Bessent's involvement and under review by White House Chief of Staff Susie Wiles, but the proposal is not public and had not been reviewed by President Trump at the time of reporting. Bloomberg also reports that U.S. export controls prompted Anthropic to temporarily disable its Fable 5 and Mythos 5 models and that OpenAI made significant changes at the government's request before releasing Sol. The White House said on July 8 that release timing remains a company decision and that no permission is required, while Axios reported that the Commerce Department's Center for AI Standards and Innovation had reviewed GPT-5.6; that center is under an acting director, NIST Director Arvind Raman, following Chris Fall's resignation confirmed on July 20.
Latest coverage
Showing newest first with full-topic pagination
Alibaba released Qwen3.8-Max through its hosted API and QwenWork on August 3, while saying the 2.4-trillion-parameter model's open weights will follow next week. The company describes it as its most capable Qwen model; independent reporting found strong Arena placements but also showed that several Anthropic models ranked higher, so Alibaba's benchmark comparisons remain vendor-reported.
OpenAI said August 1 that an internal version of its unreleased Astra model produced ten results across mathematics and theoretical computer science, including three resolutions of Erdős problems. Humans prepared the arguments as manuscripts and the model formalized them in Lean, but independent expert review will determine the correctness, novelty and significance of each claimed advance.
China's National Development and Reform Commission said on July 31 that AI-related sectors maintained growth above 30% in the first half of 2026. The agency also reported that intelligent computing capacity reached 2.8 times its year-earlier level by the end of June and outlined plans for more application pilots, data sharing and AI governance.
Moonshot AI released the downloadable weights for Kimi K3 on July 27, after unveiling its 2.8 trillion-parameter model earlier in July. Moonshot's July 16 technical blog describes Kimi K3 as a 2.8 trillion-parameter open model with a 1 million-token context window and capabilities aimed at reasoning, coding, and knowledge work. The release has intensified debate over open model diffusion, export controls, and sovereign AI deployments.
A July 30 Jamestown Foundation review of Chinese research found military- and security-linked institutions using distillation, including outputs from OpenAI and Anthropic models, to train smaller systems for defense, surveillance and cyber uses. SiliconANGLE, summarizing a Reuters review of more than 80 papers and patents, reported that such models can run locally with less compute than frontier training requires.
On July 21, 404 Media reported that ISBNdb was offering bulk printed-book sourcing for AI training, including orders of up to 1 million titles and material promoted as predating widespread generative AI. By July 28, the service page had been removed; PC Gamer reported that ISBNdb said the offer was part of exploring demand and that the company had pivoted away from that direction.
OpenAI disclosed on August 1 that an internal version of Astra, described as its next major model, produced ten results in mathematics and theoretical computer science. OpenAI reported that the problems had seen no progress on their main results for at least a decade, and released human-prepared manuscripts, Lean formalizations, and model reasoning narrations for the work.
OpenAI said on July 31 that its models reach more than 1 billion active users and more than 2 million businesses. The company also detailed price cuts announced a day earlier: GPT-5.6 Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, while Terra costs $2 and $12. The usage figures are company-reported and not independently audited.
DeepSeek released DeepSeek-V4-Flash-0731 on July 31, superseding the preview model with stronger reported agentic results and native Responses API support. The company says the update changes only the Flash API, while V4 Pro and its app and web models remain unchanged. Artificial Analysis scored Flash-0731 at 50 on its Intelligence Index, 10 points above the previous V4 Flash.
GlobaLeaks published an independent AI-assisted source-code audit on July 30, 2026 that identified 29 confirmed vulnerabilities, 12 denial-of-service issues and 42 hardening recommendations. The organization said two findings were High severity, none was critical, and all identified issues had been addressed. ISGroup reported about $3,140 in model API costs before human validation.
Sarvam AI said at its Epoch conference in Bengaluru on July 30 that it plans to build a trillion-parameter model in India for coding, cybersecurity, scientific research and simulation. The Times of India reported the roadmap directly from the event; a separate Free Press Journal report said the company expects the model within six months. No architecture, training-compute plan or independent evaluation has been published.
LG AI Research released the open-weight K-EXAONE 2.0 language model on Hugging Face on July 31, with 750 billion total parameters and 37 billion active parameters. The Apache 2.0-licensed model supports 10 languages and uses a sparse architecture; LG's model card reports 3-5x inference acceleration through MTP and DSpark speculative decoding.
OpenAI cut API prices for GPT-5.6 Terra by 20% and GPT-5.6 Luna by 80% on July 30, while also reducing how their use counts against ChatGPT Work and Codex subscription limits. OpenAI lists the new API prices at $2 per million input tokens and $12 per million output tokens for Terra, and $0.20 and $1.20, respectively, for Luna. The company also replaced Priority Processing with Fast mode for GPT-5.6 Sol.
NightRun is an experimental open-source UEFI application that boots a Raspberry Pi 5 or x86_64 PC directly into a local LLM without loading a conventional operating system. Published on GitHub on July 27 and independently described by CNX Software on July 30, the Rust runtime loads quantized 1B to 4B models into RAM and provides an offline framebuffer chat interface from USB or microSD boot media.
Anthropic resolved a Claude outage on July 29 after elevated errors affected multiple models, Claude.ai, Claude Code, and the Claude API. The official public incident record shows errors beginning at 19:45 UTC, recovery across most models later in the disruption, and resolution at 22:36 UTC.
OpenAI said on July 29 that GPT-5.6 Sol helped reduce end-to-end model-serving costs by 20% through kernel optimization and improve token-generation efficiency by more than 15% through speculative-decoding work. The figures are company-reported, and OpenAI did not announce a price cut; the practical signal is that a frontier model is now helping optimize the infrastructure that runs it.
A Communications Psychology study published April 28 analyzed 3,366 dream and waking-experience reports from 207 adults, plus 351 dream reports from 80 adults during Italy's first COVID-19 lockdown. Using LLM-assisted ratings and lexical analysis, the researchers found associations between dream content, stable personal traits, and shared external events, while emphasizing that the observational results do not establish causation.
HomomorphicEncryption.org's FHE Benchmarking Suite added a BERT workload whose current reference submission follows DESILO's THOR method, the suite's public repository showed on July 24. The workload runs BERT-Base on MRPC under encrypted inference and supplies a common harness for comparing implementations. It is a reference workload, not evidence that all generative or decoder-only language models have comparable encrypted-inference performance.
Researchers affiliated with IIIT Hyderabad and partner institutions reported that 3-billion-parameter language models matched 7B–14B models when predicting fMRI responses during naturalistic language comprehension, while 1B–1.5B models degraded. The preprint was posted on February 7; its findings suggest compact models may be adequate research probes, but they do not show that the models reproduce human language understanding.
Moonshot AI released Kimi K3's full weights on July 27, making its 2.8 trillion-parameter multimodal mixture-of-experts model available under the Kimi K3 License. The official model card lists 104 billion activated parameters, 896 experts with 16 selected per token and a 1,048,576-token context window; independent analysis puts the Hugging Face repository at roughly 1.56 TB.
Arena has launched a factuality leaderboard that ranks models on the correctness of their claims alongside human preference, built on more than 2 million labeled claims from real conversations. In an email interview with Lets Data Science, Arena said that while human preference scores have climbed across nearly every provider, OpenAI is the only provider also consistently improving factuality over an extended period, and that most open-source models trend the other way once factual accuracy is weighted in.
Singapore startup Acrab announced its GELIX 1 edge AI system-on-chip and Agent Box reference system on July 23, targeting local inference for models in the 100-billion-parameter class. Acrab says the 5 nm chip combines a 20-core Arm CPU and multicore NPU with 273 GB/s unified-memory bandwidth; independent reports describe its 7.5x M4 Pro result as an unverified company benchmark.
Vectoral published a June 28 investigation mapping how Chinese “relay” services pool and resell access to U.S. AI models, sometimes at discounts as high as 97.8% from official list prices. The report traces a four-layer supply chain and documents abuse methods including free-trial farming, chargebacks, stolen cards and exposed application endpoints. It recommends layered controls while noting that relay software itself can be legitimate.
Thinking Machines Lab released Inkling on July 15, 2026, its first general-purpose model family with downloadable weights under Apache 2.0. The company says the main mixture-of-experts model has 975 billion total parameters, 41 billion active per token, multimodal inputs, and a one-million-token context window; independent performance testing is still needed.
In a May 30 build log, Oscar Molnar described adapting a used 16GB Tesla V100 SXM2 to run beside an RTX 4080, producing 32GB of combined VRAM for about £200. The mixed-GPU system reportedly ran a 27-billion-parameter model at 32 tokens per second, though it required legacy drivers, an adapter, and custom fan control.
Anthropic released Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output tokens, the same base pricing as Opus 4.8 and half the price of Claude Fable 5. The company says Opus 5 approaches Fable 5 on many coding and knowledge-work tasks, while Axios reports it is becoming the default model for Claude Max subscribers.
Alphabet CEO Sundar Pichai said on the July 22 Q2 earnings call that coding and agentic coding are areas Google needs to improve. Google says Gemini 4 has entered its most ambitious pretraining run, and Pichai said the larger base model will be needed to compete at the next frontier. No release date, API specifications or independently verified coding results are available.
University of Nebraska-Lincoln researchers posted a July 14 revision of MechAInistic, a preprint system that turns natural-language questions into executable workflows over genome-scale metabolic models. The reviewer-supervised design was evaluated on two paired immune-cell cases, where it generated drug hypotheses for rheumatoid arthritis and multiple sclerosis. The findings remain preprint evidence and are not clinical validation.
A BlueDot sprint project found that 4-bit quantization and moderate magnitude pruning partially restored the probability of forgotten answers in one Llama-3.2-1B unlearning setup. The strongest effect—NPO with 20% pruning—recovered 42% of the gap toward the full-knowledge model, but did not restore verbatim facts. Because every cell was run once on one model and benchmark split, the result is a reproducible warning, not proof that compression broadly defeats unlearning.
Chris Fall resigned as director of the U.S. Center for AI Standards and Innovation after about three months in the role, the Commerce Department confirmed on July 20. NIST Director Arvind Raman will serve as acting director while the department searches for a permanent replacement, leaving a leadership transition at the federal center responsible for voluntary advanced-model testing and AI security evaluations.
