Model Pricing

vLLM announces day-0 support for Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with 1M-token context. Achieves up to 370 tok/s with DSpark speculative decoding. Features hybrid prefix caching, tool calling, and optimizations for production deployment.

This paper introduces Ordered Action Tokenization (OAT), a learned action tokenizer that maps continuous robot action chunks to an ordered sequence of discrete tokens, achieving high compression, total decodability, and ordered token space. It demonstrates strong performance across various tasks and backbones.

Moonshot has released Kimi K3, a 2.8 trillion parameter multimodal model, now available on Modal at 460 tokens per second. The model features a mixture-of-experts architecture, 1M token context window, native vision, and is optimized with a custom DFlash speculator for faster inference.

Matt Lenhard's investigation reveals a market where LLM tokens are resold at a discount by abusing free trials, unprotected support bots, stolen credit cards, and chargeback attacks. Primarily in China, resellers use open-source proxy software like one-api and new-api to pool API keys. Buyers seek cheap tokens, avoid geo-restrictions, or collect data for model distillation. The author calls for stricter API key caps from vendors.

A developer successfully runs a 28.9 million parameter language model on an ESP32-S3 microcontroller costing around $8, using Per-Layer Embeddings to store most parameters in flash. The text generation runs entirely on-device at about 9.5 tokens per second, a significant leap from previous 260k-parameter models.

Anthropic has released Claude Opus 5, replacing Opus 4.8 as the Opus-tier flagship. Pricing remains unchanged at $5/M input and $25/M output tokens. The model approaches Claude Fable 5's intelligence at half the price. Key API changes: thinking enabled by default, breaking change for disabled thinking at high effort, and removal of verification prompts. It achieves strong results in agentic coding benchmarks like OSWorld (70.57%) and Zapier AutomationBench (26.0%), as well as reasoning tests (perfect score on IMO 2026 problems, 30.16% on ARC-AGI-3). Safety relaxations are limited to source-code vulnerability finding; exploitation remains blocked.

Kolega Code is a local-first terminal coding agent with multi-agent orchestration (Gigacode) for broad tasks like large audits, migrations, and parallel checks. It supports model routing, plan/build modes, web search, MCP servers, and is open source under Apache 2.0.

cloro.dev, the leading AI UI scraping platform, introduces a new recurring free tier offering 500 credits every month. The platform extracts structured data from ChatGPT, Perplexity, Grok, Gemini, Google Search, Google News, Copilot, and AI Overview via a single API. It has crossed 200 customers, mostly large enterprises.

A new study reveals that Mixture-of-Experts (MoE) routing in large language models follows a principle similar to Huffman coding, where common tokens are processed by sparse experts and rare, complex tasks engage diverse expert committees. The paper introduces the Frequency-Diversity Law, identifies a redundancy trap in certain models, and proposes Subset Difference Pruning to improve efficiency.

Kolega Code is an open-source, local-first CLI tool that orchestrates multiple AI agents for coding tasks. It supports various model providers, features parallel sub-agent workflows (Gigacode), web search, browser automation, and keeps all data on the user's machine.

This paper introduces ChronoStitch, a training-free method for composing independently stored visual key-value (KV) memories to enable long-horizon temporal reasoning in video question answering. By re-basing stored post-rotary keys onto a global three-axis multimodal RoPE coordinate system and selectively recomputing high-deviation visual tokens, it overcomes temporal phase collisions and content gaps from naive concatenation. Experiments on Qwen2.5-VL-3B and the temporal split of TempCompass show improved event-ordering accuracy and 3.3x speedup over full joint re-prefilling.

LiquidBrain.ai offers unlimited tokens and unlimited context at a fixed price, with a strong emphasis on data privacy.

On July 21, 2026, Google quietly released Gemini 3.6 Flash, a mid-cycle update focused on efficiency rather than breakthrough capability. It maintains similar reasoning to 3.5 Flash but with significantly reduced token usage and cost. Improvements in coding, ML tasks, and computer use are notable, with a refreshed knowledge cutoff. The model is priced at $1.50/M input tokens and $7.50/M output, cheaper than its predecessor. The article includes stress tests for readers to evaluate the model themselves.

Thinking Machine's new model Inkling has 975 billion parameters but activates only 41 billion per token, using a router to select specialist subsets for sparse computation.

The US Army exhausted its entire year's supply of AI tokens for Ask Sage in just one month, forcing a reimposition of limits. Despite encouragement to use generative AI heavily, token consumption has been staggering, raising questions about the utility and reliability of AI tools.

Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber on July 21, 2026. The Flash tier gets cheaper and more token-efficient, with 3.6 Flash cutting output tokens 17% and dropping its output price to $7.50 per 1M. Flash-Lite runs at 350 tokens/sec, while gated Flash Cyber powers CodeMender for vulnerability finding. The flagship 3.5 Pro remains delayed.

This post explores generating thinking tokens for datasets lacking reasoning traces in SFT customization. It examines the reasoning suppression problem, introduces Self-Distilled Reasoning (SDR), validates it across three benchmarks, and provides practical recommendations. SDR reuses the base model's chain of thought as a stand-in, mitigating catastrophic forgetting while maintaining or improving target performance.

Calyxa is a browser-native AI tutor that helps students learn directly on homework websites and PDFs, eliminating the need to copy-paste problems into separate chats. It provides contextual annotations and voice coaching, aiming to turn cheating into genuine understanding. Free tier offers 10 sessions/month; Pro at $10/month for unlimited use.

Mixture-of-Experts (MoE) scales Transformers by routing tokens to a subset of experts, but existing routers use shallow or isolated token representations, leading to unstable and semantically inconsistent routing. This work proposes Multi-level Context Fusion MOE (MCF-MOE), which integrates cross-layer semantic aggregation and local token-level interactions for more context-aware representations. Experiments show improved routing consistency and downstream performance.

A new reinforcement learning method, W2SPO, uses weak auxiliary models to inject short segments into LLM reasoning paths, improving performance and training speed on math reasoning tasks.

Portraify is an AI-powered headshot generator that lets users upload 1–3 everyday photos to get a studio-quality portrait in seconds. It offers a free tier with 3 portraits, paid plans starting at $9, and emphasizes privacy by not storing uploaded photos.

Italian engineer Vincenzo (JustVugg) created Colibrì, a proof-of-concept that runs the 744-billion-parameter GLM-5.2 model (1.5TB) on a modest CPU with only 25GB RAM and 1GB/s NVMe. Despite extremely slow speeds (0.05-0.1 tokens per second), it leverages the Mixture-of-Experts architecture to load experts per token, achieving frontier-level answer quality. The project is open-source and aims to explore running large models on consumer hardware.

Thinking Machines Lab has released Inkling, its first general-purpose open-weights foundation model. It is a multimodal MoE model with 975B parameters, 41B active parameters, and a 1M-token context window. Designed for customization, Inkling excels in reasoning, coding, agentic workflows, and multimodal tasks. This guide covers its architecture, training, benchmarks, deployment, and fine-tuning workflow.

The paper introduces BIRD, a two-stage self-reasoning distillation method that first samples concise solutions with a brevity instruction and performs prompt-switch SFT, then applies on-policy reverse-KL distillation on cleaner prefixes. On Qwen3-8B, MATH-500 accuracy improves from 86.2% to 92.0% while response length drops from 3,099 to 1,115 tokens.

A systematic evaluation of five major LLMs for technical market analysis finds GPT-4 Turbo achieves highest annualized return and Sharpe ratio, while FinGPT shows competitive risk-adjusted performance through domain fine-tuning. The study also identifies failure modes including numerical hallucination and context window limitations.

Apple researchers present the Length Value Model (LenVM), a token-level framework that predicts remaining generation length at each decoding step. By framing length modeling as a value estimation problem with a constant negative reward per token, LenVM provides annotation-free, dense, unbiased, and scalable supervision. Experiments on LLMs and VLMs show that on the LIFEBench exact length matching task, LenVM improves a 7B model's length score from 30.9 to 64.8, surpassing frontier closed-source models. LenVM also enables continuous control over the performance-efficiency trade-off and predicts total generation length from the prompt.

Introducing a free tool to analyze token usage in Fable's AI harness.

This article explores how to develop reasoning models with multiple effort modes, covering the evolution from o1 and DeepSeek-R1 to GPT-5.6, and key techniques such as RLVR training, inference scaling, think tokens, and reasoning mode toggles.

Lowest cost per token from extreme codesign maximizes intelligence per dollar for post-training in the agentic era.

Polestar is a training-free inference framework that addresses KV-cache reuse and decoding parallelism challenges in diffusion LLMs by leveraging token representation drift. It consists of Polestar-Cache for sparse cache refreshes and Polestar-Commit for identifying commit-ready tokens, achieving up to 10.73% accuracy improvement and 3.7x higher throughput on math and coding benchmarks.

This paper introduces token time continuous diffusion (TTCD), a diffusion language model operating in continuous space with per-token times, where tokens proceed from noise to token at varying rates. TTCD avoids parallel sampling inaccuracies and outperforms discrete models at high speedups. A 160M parameter model trained on OpenWebText and self-distilled achieves comparable unconditional and superior conditional generation, with gains in Sudoku solving.

Puter compiled Firefox to WebAssembly, enabling a full browser to run inside another browser. The project used an estimated $25,000 in Claude Opus and Fable tokens, leverages the Wisp protocol for proxying, supports end-to-end encryption, and is open source.

xAI's Grok 4.3 is now generally available on Amazon Bedrock, offering configurable reasoning effort, strong tool use, instruction following, and a 1 million token context window for agentic and enterprise workloads. This post covers its features, access methods, and how to use key capabilities such as chat, reasoning, tool calling, structured output, image input, and multi-turn conversations.

Mira Murati's Thinking Machines Lab released Inkling, a 975B parameter MoE model (41B active) under Apache-2.0 license, multimodal, trained on 45T tokens. It's not frontier but a strong base for fine-tuning via Tinker platform. Inkling-Small (276B, 12B active) is promised. Model card and training data documentation are unusually brief. Inkling is competitive with Chinese open-weight models, adding to the US ecosystem.

A technical report from arXiv introduces Oracle Agent Memory, a database-native memory system built on Oracle Database for long-horizon AI agents. It achieves 93.8% accuracy on LongMemEval while using 10.7x fewer tokens compared to flat-history baselines. The system addresses memory lifecycle, layered architecture with scope control, and evaluation methodology combining task accuracy with memory-specific metrics.

Voicebox is an open-source voice-to-text tool that captures speech, transcribes it via Whisper, and formats the output with an LLM, all powered by Cloudflare Workers AI.

Throttle is a macOS menu bar meter for Claude Code usage that evolved into a full cockpit. The free version provides local, no-telemetry monitoring. Pro adds a project cockpit with embedded terminals, auto-hibernate, remote session transfer to Linux servers, and an AI optimizer that audits claude.md and settings.json to reduce output tokens by 65–75%. All data stays local or in iCloud private database. One-time fee of €29.

Mira Murati's Thinking Machines Lab Inc. today launched its first foundation model with the release of Inkling, making its full open weights available to developers so they can fine-tune it as they wish. Inkling is a mixture-of-experts model with 975 billion parameters (41B active) trained on 45 trillion tokens of text, image, audio and video, capable of reasoning across all four modalities but outputting only text. It features "thinking effort" controls and uncertainty flagging to reduce hallucinations. The model is fine-tunable via the Tinker API and aims to provide a Western open-source alternative to Chinese AI models. Thinking Machines plans to generate revenue through the Tinker platform rather than per-token API access, potentially disrupting current AI business models.

Thinking Machines Lab released Inkling on July 15, 2026, its first model trained from scratch. The full weights ship under Apache 2.0. It is a 975B-parameter Mixture-of-Experts transformer with 41B active parameters, a 1M-token context window, and native text, image, and audio input. The core differentiator is controllable thinking effort, allowing users to adjust token budgets per call to balance cost and performance.

A German research consortium has published the pretraining report for Soofi S 30B-A3B, an open base model for German and English. It is a Mixture-of-Experts hybrid Mamba Transformer model with 31.6B total parameters, activating 3.2B per token. It achieves the highest English and German aggregate scores among tested fully open base models.

Limits is an iOS app that monitors AI tool quotas for Codex, Claude Code, and Cursor directly on your device. It sends notifications when limits reset, predicts when you'll run out, and helps you redeem expiring rate-limit resets. All data stays on your phone, with tokens stored in the iOS Keychain.

This paper shows that scaling can substantially narrow the performance gap between point-in-time language models and their unconstrained counterparts. The authors trained decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens from FineWeb, creating monthly model checkpoints from 2013 to 2024. On reasoning and understanding benchmarks, these models approach the performance of leading open-weight models of comparable size (e.g., Gemma-3-4B and LLaMA-7B) trained on temporally unrestricted data. Instruction fine-tuning via LoRA further improves downstream usability. The complete pipeline is released for reproducibility.

PrismML just released Bonsai 27B. It is a low-bit representation of Qwen3.6-27B, not a new pretrain. The architecture is unchanged. Two variants ship under Apache 2.0. Ternary Bonsai 27B uses {−1, 0, +1} weights at a true 1.71 bits per weight. Its ideal size is 5.9GB. 1-bit Bonsai 27B uses binary {−1, +1} weights at 1.125 bits per weight, for 3.9GB. Performance: ternary retains 94.6% of FP16, binary retains 89.5%. Both are multimodal, context 262K tokens. PrismML claims the 1-bit build is the first 27B-class model to fit a phone.

A new study reveals that coding agents need minimal context when editing code: the signal is only in the code being edited, natural-language summaries fail to answer behavioral questions, surrounding context (UML skeletons) performs no better than deleting it, and compressed context matches full files at one-third the tokens. Temperature-0 inference introduces a ~9% noise floor. The authors release their instrument including gold-validated environments, deterministic patches, and pre-registered hypotheses.

Anthropic has launched Claude Sonnet 5, its most agentic mid-tier model, outperforming Sonnet 4.6 across all benchmarks and narrowing the gap to Opus 4.8. It introduces effort levels to control reasoning costs, offering great value at low/medium effort but potentially exceeding Opus 4.8 cost at extra-high effort. It is now the default model for Free and Pro plans and accessible via API.

QuantumReckon is a new tool that reveals the true cost of cloud and AI spending, especially the often-hidden token costs from AI APIs. It connects to multiple cloud and AI providers, performs daily automated sweeps, detects anomalies and waste, and provides auditable evidence with sealed receipts. Validated on the founder's own estate, it identified significant savings.

CLAP converts pretrained VLMs to VLAs by prepending language descriptions to action tokens, avoiding distribution shift. Single-epoch fine-tuning yields 90.8% on LIBERO (+14.9 over VLA-0) and improved robustness. Open-weight models at 0.8B, 2B, 4B to be released.

We propose StickyMoE, a differentiable routing consistency loss that penalizes abrupt expert switches between adjacent tokens during training, enabling memory-efficient inference on edge devices. Experiments show up to 60% reduction in expert switch rate with less than 4% perplexity degradation.

This paper identifies a failure mode called Positive-Credit Contamination in RL for LLMs, where low-probability erroneous tokens receive identical positive credit as plausible ones. The proposed TACO method computes a tail-risk score to calibrate credit assignment, outperforming GRPO baselines across three LLMs and eight benchmarks while improving training stability in long-horizon RL.

Jet-Long introduces a tuning-free zero-shot method for extending LLM context windows by using dynamic bifocal RoPE, which adapts the rescaling factor to sequence length, achieving high efficiency and strong performance on multiple benchmarks.