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 6, 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 results depend on the harness as much as the checkpoint and models that optimize against the scoring mechanism 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 open-weight tier and the frontier moved at the same time, and they are now measured in different currencies. Alibaba opened access to Qwen3.8-Max through its hosted API and QwenWork on August 3, describing a mixture-of-experts model with 2.4 trillion total and 95 billion active parameters and saying downloadable weights would follow the next week; independent reporting found strong Arena placements but also showed several Anthropic models ranking higher, so its own benchmark comparisons remain vendor-reported. Moonshot AI published Kimi K3's full weights on July 27 under the Kimi K3 License, 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, and an independent estimate putting 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 with a one-million-token context window, and LG AI Research published K-EXAONE 2.0 on July 31 at 750 billion total and 37 billion active, also Apache 2.0, supporting 10 languages with a reported 3 to 5 times inference acceleration from multi-token prediction and DSpark speculative decoding. At the other end, OpenAI disclosed on August 1 that an internal version of Astra, which it describes as its next major model, produced ten results across mathematics and theoretical computer science, including three resolutions of Erdős problems on which OpenAI says there had been no progress on the main results for at least a decade. Humans prepared the arguments as manuscripts and the model formalized each in Lean, so the release shipped inspectable proof artifacts rather than benchmark scores, but Astra is unreleased, OpenAI published no complete success rate across attempted problems, and independent expert review will decide correctness, novelty and significance.
Pricing and verification moved in opposite directions in the same window. 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 Axios reported it becoming 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, reduced how their use counts against ChatGPT Work and Codex subscription limits, and replaced Priority Processing with Fast mode for Sol; it said on July 29 that Sol-assisted kernel work had cut end-to-end serving costs by 20% and that a Sol-improved speculative-decoding draft model raised token-generation efficiency by more than 15%, both company-reported with no customer price change attached, and said on July 31 that its models reach more than 1 billion active users and more than 2 million businesses on figures that are not independently audited. 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 case: Artificial Analysis independently scored it 50 on its Intelligence Index, 10 points above the previous V4 Flash, while DeepSeek's own coding-agent gains rest partly on unreleased evaluation infrastructure and internal test sets. Lasso Security showed on August 3 that holding everything else fixed and switching only the agent harness flipped 43 of 100 model-and-mission pairings from at least one success to none across a 1,000-attack run, and its independent judge rejected 155 of 303 attacker-declared wins. ProjectDiscovery's audit of 54 usable black-box web targets found most failures came after the correct vulnerability had already been identified. Arena, whose factuality leaderboard rests on more than 2 million labeled claims from real conversations, 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 weighted in. Policy tightened around the same models: European Commission enforcement powers over general-purpose AI providers took effect on August 2, allowing information requests, model evaluations, corrective measures or market withdrawal and fines up to 15 million euros or 3% of worldwide annual turnover, while a July 30 Jamestown Foundation review found Chinese military- and security-linked institutions using distillation, including outputs from OpenAI and Anthropic models, to train smaller systems for defense, surveillance and cyber work. For anyone building or buying, the 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. Alibaba said on August 3 that Qwen3.8-Max open weights would follow the next week, so that is directly checkable, as is whether any third party reproduces Moonshot's Kimi K3 benchmark table, which coverage notes mixes company-run tests, public leaderboard results and different agent harnesses. DeepSeek's published code-agent scores rest on evaluation infrastructure and internal test sets that have not been released. OpenAI has published no model card, architecture, context-window specification, API details or release date for Astra, and has not published a success rate across attempted problems, so the meaningful signal is whether independent mathematicians inspect the Lean formalizations. Google says Gemini 4 has entered its most ambitious pretraining run with no release date, parameter count, context window, pricing or reproducible coding evaluations disclosed, and Sundar Pichai said on the July 22 earnings call that coding and agentic coding are areas Google needs to improve. 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, no binding U.S. regime is established: Bloomberg reports the proposed FINRA-like AI regulator reporting to the SEC remains under deliberation and had not been reviewed by President Trump, and Chris Fall resigned as director of the U.S. Center for AI Standards and Innovation after about three months, with NIST Director Arvind Raman serving as acting director. In the EU, the question is whether the Commission's new powers over general-purpose AI providers produce any information request, evaluation or corrective measure in practice. Finally, DeepSeek's reported discussions about a second private financing at roughly a $71 billion pre-money valuation had not closed and the company has not publicly confirmed them.
It is possible, and the cost moves onto you. Moonshot's Kimi K3 lists 2.8 trillion total parameters with 104 billion activated and a 1,048,576-token context window, with the Hugging Face repository estimated at roughly 1.56 TB; Nature reported that size alone could limit adoption even if capabilities hold up. Lighter options exist at the same tier: Thinking Machines Lab's Inkling is 975 billion total and 41 billion active under Apache 2.0, and LG's K-EXAONE 2.0 is 750 billion total and 37 billion active, also Apache 2.0. Alibaba said Qwen3.8-Max weights, at 2.4 trillion total and 95 billion active, would follow its August 3 hosted release. Downloading weights shifts infrastructure, evaluation, security review and monitoring costs to whoever runs them.
Meaningfully at the commodity tier and less at the top. Anthropic released Claude Opus 5 on July 24 at $5 per million input and $25 per million output tokens, unchanged from Opus 4.8's base pricing and half the price of Claude Fable 5. OpenAI cut GPT-5.6 Terra by 20% to $2 and $12 and Luna by 80% to $0.20 and $1.20 on July 30, and also reduced how their use counts against ChatGPT Work and Codex subscription limits. Separately OpenAI said Sol-assisted kernel optimization cut its own end-to-end serving costs by 20% with no announced customer price change, which is a reminder that provider cost and list price are different numbers.
Treat it as a claim about a system, not a model. Lasso Security held the model, prompt, tools, targets, gateway and judge fixed and changed only the harness across a 1,000-attack run; average objective success stayed close at about 21% and 19%, yet 43 of 100 model-and-mission pairings flipped from at least one success to none, and an independent judge rejected 155 of 303 attacker-declared wins. ProjectDiscovery found most failures on 54 black-box web targets came after the right vulnerability had been identified, an execution gap solve rates hide. Coverage of Moonshot's Kimi K3 benchmark table notes it mixes company-run tests, public leaderboard results and different agent harnesses. Reproduce inside your own harness before you plan around a number.
Partly, which is the point of the release format. OpenAI said on August 1 that an internal Astra version produced ten results in mathematics and theoretical computer science, including three resolutions of Erdős problems, and published human-prepared manuscripts, a Lean formalization of each argument and model narrations of the reasoning. Lean formalizations are inspectable proof artifacts, so correctness is verifiable in a way benchmark scores are not. Novelty and significance still require expert review, and OpenAI published no complete success rate across attempted problems, no model card, no architecture, no context-window specification, no API details and no release date.
European Commission enforcement powers over general-purpose AI providers took effect on August 2, allowing the Commission to request information, evaluate models, order corrective measures or market withdrawal and impose fines up to 15 million euros or 3% of worldwide annual turnover; that makes duties applying since August 2025 enforceable without establishing that any provider has breached the Act. A July 30 Jamestown Foundation review found Chinese military- and security-linked institutions using distillation, including outputs from OpenAI and Anthropic models, to train smaller systems for defense, surveillance and cyber uses, based on documented research and proposed uses rather than operational deployment. In the United States no binding regime exists: a proposed FINRA-like regulator reporting to the SEC remains under deliberation, and the Center for AI Standards and Innovation is being run by an acting director after Chris Fall's July resignation.
Smaller than the frontier discussion suggests, with real tradeoffs. NightRun boots a Raspberry Pi 5 or x86_64 PC directly into a quantized 1B to 4B model from UEFI with no conventional operating system, using CPU vector kernels with memory bandwidth as the main decoding constraint and giving up networking, observability and package management in exchange. Acrab's GELIX 1 edge SoC targets local inference for models in the 100-billion-parameter class with 273 GB/s unified-memory bandwidth, though its comparison against an M4 Pro is an unverified company benchmark. 9to5Mac, citing The Information, reported Apple interest in PrismML technology for running compressed models on-device, including a 27-billion-parameter Qwen 3.6 on an iPhone 17 Pro. A separate BlueDot result is a caution for compressed deployments: 4-bit quantization and moderate pruning partially restored the probability of forgotten answers in one Llama-3.2-1B unlearning setup.
Latest coverage
Showing newest first with full-topic pagination
NokiaPowerUser reported on August 23 that Gemini 3.5 Pro could launch that day, citing leaked specifications including a possible 2 million-token context window. Google has not announced a broad release: its July 21 blog post stated that Gemini 3.5 Pro was testing with partners and would become broadly available when ready.
Simon Willison released LLM 0.33 on August 22, moving the project's HTTP dependency from httpx to httpx2 after a short-term compatibility fix in the prior release. The official GitHub release is a non-prerelease version published at 17:01 UTC, making this a developer-environment update rather than a model announcement.
An anonymous model called Ox Alpha appeared on OpenRouter on August 20 with a 1,048,576-token context window and text, image, and video inputs, according to OpenRouter listing details reported by GLM5.app. Community testing has linked its tokenizer, video behavior, and backend errors to Zhipu AI's GLM family, but neither Zhipu AI nor OpenRouter has publicly claimed the model. Early coding benchmark claims remain unverified.
An August 11 preprint analyzing 1.19 million open-access biomedical papers in PubMed Central estimated that 89% of papers published in December 2025 showed an excess of vocabulary associated with large-language-model assistance. The authors describe an estimate of assisted-writing prevalence, not proof that AI generated an entire paper, with implications for publication-policy debates.
Hebrew University researchers reported in iScience on August 6 that GPT-4 generated personality questionnaires from DSM-5 and astrology texts, then forecast population-level response patterns before 600 participants completed them. The DSM-based questionnaire showed stronger internal consistency than the astrology-based version, a reminder that generated survey items still require human validation.
OpenAI cut developer pricing for GPT-5.6 Sol by more than 20% for the next three months, Reuters reported on August 21. Standard short-context API pricing fell to $4 per million input tokens and $20 per million output tokens, from $5 and $30, respectively. The lower rates also apply to eligible credits for ChatGPT Work and Codex, while Pro, Plus, and Business subscription pricing is unchanged.
DeepSeek released the experimental DeepSeek-V4-Flash-Vision-Exp multimodal model on its API platform on August 21, adding image understanding to its V4-Flash model line. DeepSeek reports that the model retains V4-Flash text capabilities and brings multimodal agent results close to Anthropic's Opus-4.8 on its published benchmarks. The release also adds Files API support and DeepSeek Harness 0.1.1 compatibility.
Artificial Analysis lists Qwen3.8 27B at 52 on its Intelligence Index, placing the 27-billion-parameter open-weights model among the stronger small models in its August comparison. The benchmark page also lists a 256,000-token context window. The score is a comparative benchmark result, not a guarantee of equivalent performance on every production workload.
An Aug. 19 research preprint introduced LeakGauge, a method that uses prefill token probabilities to estimate whether an LLM prompt may trigger context leakage before generation begins. Across 11 tested models, the authors report AUROC scores of 0.944 to 0.996 and released code for the approach; the findings have not yet been independently replicated.
Google rolled out Gemini 3.7 Flash to Google AI Pro and AI Ultra subscribers on August 20, according to NokiaPowerUser. Google introduced the model on August 13 for developer channels, describing it as a multimodal reasoning model for coding and agentic workloads. Official documentation lists a 1 million-token input limit, 65,536-token output limit, and configurable thinking settings.
Alibaba released Qwen3.8-27B, a 27-billion-parameter multimodal model whose official repository uses Apache 2.0, and published weights for the 2.4-trillion-parameter Qwen3.8-2.4T-A95B model. CNBC reported the rollout on August 17 and Alibaba's claim that the 27B model can match a model ten times larger. The two releases have different license terms, so weight availability does not imply identical commercial rights.
Google Research published findings on August 12 showing that factual errors in frontier LLMs can stem from recall failures rather than missing encoded knowledge. Search Engine Journal reports that the evaluated models encoded 95-98% of tested facts but could not directly recall 26-34%, with reverse subject-object question phrasing contributing to retrieval difficulty.
Z.ai unveiled GLM-5.3 on August 14, reporting large post-training gains for long-horizon coding and cybersecurity tasks without changing GLM-5.2's 743 billion-parameter base model. The company reports an 84.5% CyberGym score but a lower 54.4% ExploitBench result, and says open weights will follow a two-week safety review and hardening period.
Alibaba released Qwen3.8-2.4T-A95B, the text-only open-weight checkpoint behind its managed Qwen3.8-Max service, on August 12. The mixture-of-experts model has 2.4 trillion total parameters, activates 95 billion per token, and has a 262,144-token native context extendable to 1,010,000. Its custom license requires a separate agreement for certain model-service or AI-work-assistant businesses once aggregate 12-month revenue exceeds $50 million.
Debian developers opened a ranked General Resolution vote on August 15 to determine how large language model-assisted contributions may be handled in the Linux distribution. According to Phoronix and Linuxiac, the ballot contains eight proposals, from a Social Contract ban to conditional acceptance, plus a "None of the above" option; voting closes August 28.
Google began rolling out Gemini 3.7 Flash as a selectable model in Search's AI Mode on August 14. The option is available globally in English to Google AI Pro and Ultra subscribers, but Google has not said it will replace AI Mode's default model or become available to the free tier.
Alibaba's Qwen team released Qwen3.8-27B on August 14 under Apache 2.0, offering a dense 27-billion-parameter vision-language model with a 262,144-token native context window and downloadable weights. Qwen's own results put it above Claude Opus 4.6 Max on SWE-bench Pro but below it on several other tests, broadening local deployment options without establishing across-the-board frontier-model equivalence.
WRITER released Palmyra X6, a new flagship model, alongside a rebuilt Agent harness and governance tools on August 13. The company reports that the combined model and harness reduce average agent costs by 52%, improve speed by 48%, and raise quality by 10% versus prior performance. The release targets enterprise teams managing token spend across multi-step AI workflows.
Simon Willison released llm-gemini 0.33 on August 13, adding Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and two embedding models to the LLM command-line plugin. The update also requires LLM 0.32 and moves Google Search, URL context, and code execution onto LLM's typed server-side tool interface while preserving Gemini thought signatures across tool-use turns.
Google released Gemini 3.7 Flash on August 13, 2026, a multimodal model for coding and agent workflows. Google reports higher results than Gemini 3.6 Flash on its cited software engineering, web development, document-processing, and workflow benchmarks. The model supports a 1 million-token context window and is available through Google's developer and enterprise channels, with introductory API pricing through December 31, 2026.
Mother Jones reporter Matt Smith said on August 13 that ChatGPT, Claude, Gemini and Perplexity supplied parts lists, assembly guidance and code during an 11-month attempt to build an autonomous attack drone from commodity components. The prototype eventually flew a GPS-guided route and crashed after an altitude-programming failure; Smith did not complete a weapon that could autonomously select and strike a target.
A Bitcoin Policy Institute open letter published August 10 asks frontier AI labs to give vetted security researchers access to their strongest cyber-capable models before broad release. More than 40 Bitcoin and crypto organizations backed the request, arguing that current restrictions can leave defenders with weaker tools than attackers may obtain. Coinbase has received access to Anthropic's restricted Mythos model, according to Cointelegraph.
Just Facts tested paid versions of ChatGPT, Gemini, Grok and Claude with 100 political questions designed to elicit falsehoods from both ideological directions in a study reported by Fox News. The study found ChatGPT, Gemini and Claude scored lower on prompts tied to left-coded falsehoods, while Grok showed the reverse pattern and roughly half of cited AI sources were illegitimate.
Ai2, the nonprofit lab behind the Olmo models, answered Lets Data Science's questions in writing about what releasing training data, code and intermediate checkpoints enables that open weights alone cannot. The lab says the evidence for a whole class of research questions sits upstream of the final checkpoint, points to outside work hunting paraphrased benchmark contamination in its public training data, and argues the remaining open-closed gap is increasingly about surrounding infrastructure rather than the models. It also confirmed mixture-of-experts models in the next Olmo generation.
xAI released Grok 4.6 on August 12, adding a frontier model for coding, long-running agents, knowledge work, and visual project creation. xAI's developer documentation lists a 500,000-token context window and API pricing of $2 per million input tokens and $6 per million output tokens. Artificial Analysis scored the model 61 on its Intelligence Index, level with GPT-5.6 Sol in that evaluation.
DeepSeek's API documentation maps the deepseek-v4-pro endpoint to DeepSeek-V4-Pro-0813 and lists a 1 million-token context window with up to 384,000 output tokens. The page now shows time-dependent V4 Pro rates of $0.66 per million uncached input tokens off-peak and $1.32 at peak, with a 500-request concurrency limit.
Elon Musk said on July 21 that SpaceX's engineering data will be added during supplemental training for a future Grok run, excluding material restricted by US arms-export rules. The announcement describes a planned use of internal technical data, but it does not disclose the dataset's size, governance controls, training schedule, or measured effect on model performance.
36Kr reported on August 11 that ByteDance had begun consolidating AI data teams into a top-level data and security unit in early June. The unit, led by former TikTok executive Wang Yinglei, is said to cover sourcing, cleaning, quality evaluation, and security for foundation-model training; ByteDance had not publicly confirmed the reorganization when the report was published.
A University of Washington study of 23,800 English-language story completions from six large language models found that only 2% of animal characters were female, compared with 41% male and 57% neutral or ungendered. Presented at ACM FAccT on June 25, the work suggests that apparent neutrality can suppress non-masculine representation rather than correct gender bias.
A study released on August 7 compared 10 commercial language models with 50,116 respondents in the European Social Survey and found that model answers aligned unevenly across countries and socio-demographic groups. Country of residence alone explained 7.3% to 27.8% of individual alignment-score variation, depending on the model, while education, income, occupation, and religion also showed material differences.
