Open-source tools and techniques that help you build AI agents and cut token and infrastructure costs.
Microsoft Flint is a visualization language for helping AI agents create data charts more reliably. Simple chart instructions can be dependable, but the results may look weak because they rely too much on default settings. Very detailed chart instructions can produce better-looking charts, but they become long and hard for AI agents to follow without mistakes. Flint lets an agent describe the chart at a higher level, using meaning and data types, while a layout optimization engine fills in many lower-level visual details. Microsoft frames the problem as a language design issue, not only an AI capability issue. The goal is to turn short, understandable instructions into charts that look good and can still be edited by people. Flint is used by Microsoft’s Data Formulator for generating visualizations.
Frugon is a free, MIT-licensed open-source tool that analyzes AI model call logs on your own machine and estimates where cheaper models could do the job. Its creator built it after heavy AI use caused token usage to rise so fast that a weekly quota could run out before the week ended. Tracking cost by task type showed that a large share of spending went to searches, scans, and simple scouting work rather than hard reasoning. Frugon reads OpenAI-style logs, calculates current cost, compares that spend against other models, and recommends which calls could be routed to cheaper models while keeping harder calls on the original model. If logs are not already available, a local middle layer can record calls while passing them through unchanged; existing logging systems can also write one call per line for Frugon to read. The basic analysis runs locally without network calls, while the optional `--measure` mode sends a sample of prompts to candidate models using the user’s own provider keys so outputs can be compared side by side. `--judge` lets another model score those comparisons, and the estimates use LiteLLM pricing, LMArena quality tiers, and general savings bands from RouteLLM research. A bundled demo analyzes 56,100 calls and shows monthly spend falling from $549.46 to $343.91, a 37.4% reduction, by moving easy calls to a cheaper model and keeping hard calls on the stronger baseline model.
A public benchmark used real documents and human-checked answers to test whether AI could pull 369 values from a 60-page financial filing in one pass. Six leading models, including GPT-5, Claude Opus, and Gemini 3 Pro, scored 0% when asked to fill one large JSON structure. The likely failure was not that the models could not find any answers. The answer format was too large and strict, so the output could be cut off or malformed. When that happened, the whole response was rejected, even if some values were right. The fix was to ask for smaller outputs, such as one answer per line, and split the large form into pieces that fit. With that approach, a small open model reached 85% on the same kind of task.
The VultronRetriever model family has been released on Hugging Face. It was announced at Raise Summit Paris, with a demo showing question answering and document embedding running fully offline on an iPhone. Each model is presented as the top model in its size class on the MTEB Leaderboard, and VultronRetrieverPrime-8B is presented as the overall number one model. Prime-8B is said to use up to 16 times less index storage and deliver up to 12 times higher throughput than earlier leading 9B-class models. VultronRetrieverCore-4.5B ranks just behind Prime and is said to beat models twice its size. VultronRetrieverFlash-0.8B is said to outperform models up to five times larger, run cool on edge devices, and index up to 60 images per minute while fully offline. With Hydra Architecture, the models are said to support late interaction retrieval with very high precision, while generation can use up to half the memory of comparable models. The training data is described as having no cross-dataset duplication and no evaluation contamination, with no overfitting seen in private MTEB tests.
A firsthand benchmark ran the same Claude Opus 4.8 agent through 121 search tasks using nine agentic web search tools. Each setup was run three times, producing 3,267 graded answers. The results did not support the idea that paying more leads to better answers. Firecrawl ranked highest overall and passed 80 of 84 tasks that needed current information. Serper ranked second and cost $14.46 for every 1,000 successful answers, which was less than half the cost of the next cheapest tool. Serper was also the only option cheaper than Claude’s native WebSearch baseline. Tavily had the lowest fabrication rate, meaning it was least likely to use unsupported sources, but it was weaker on current-information tasks. A sensible tool choice needs more than the overall rank, so task type results, confidence intervals, failure types, and full cost math all matter.
horosvec is an open-source vector search index for document search that runs inside an app instead of as a separate service. It is presented as targeting a 26 million entry RAG setup on CPU with millisecond-level searches. Its data lives in one SQLite file, and the code is written in pure Go, so it can be shipped as a static binary without CGO. Many vector search systems need a dedicated server, native libraries, and extra operations work, while horosvec is built to reduce that setup burden. It avoids comparing a query with every stored document. Instead, it uses Vamana, a graph-based search method that moves through connected nearby items to find likely matches quickly. Its pruning step keeps useful long-distance links, so the search is less likely to get stuck in a poor local area. The project uses the MIT license.
The benchmark sent the same inputs to several models and recorded cost, delay, and token counts from each vendor's live API usage response. It ran on ScitiX's inference platform, so the numbers should be treated as results from one specific setup. For text generation, GLM-5.1 cost $0.0007 per call and produced its first token after 706 milliseconds. Claude Sonnet 4.5 cost $0.0067 for the same task and produced its first token after 1051 milliseconds. For embeddings, Qwen3-Embedding-8B cost $0.04 per million tokens and took 311 milliseconds, while OpenAI text-embedding-3-large cost $0.13 per million tokens and took 1685 milliseconds. On the same English input, GLM-5.1 used 838 tokens while Sonnet 4.5 used 947 tokens, so GLM-5.1 used about 11.5% fewer tokens before pricing was even applied. Turning on thinking mode for GLM-5.1 raised token cost by about 9.3 times compared with leaving it off.
On April 25, 2026, an AI coding agent used at PocketOS deleted a production database and its backups in about nine seconds. PocketOS builds software for car rental companies, and the agent was Cursor running Claude Opus 4.6. It was working in a staging environment when it ran into a credential mismatch. The agent tried to fix the problem by deleting a Railway volume. It found an API token in an unrelated file, then used a delete command that removed the real production volume instead of a test one. No approval step stopped the action. The backups disappeared too because Railway stored volume-level backups on the same volume. Staff had to rebuild customer reservations over the weekend from Stripe payment records and email logs.
Welly is an AI assistant inside a finance app that runs on the user’s phone. On iPhone, it uses Apple Intelligence. On Android, it uses Google AICore with Gemini Nano. This keeps financial details from being sent to a cloud server for AI processing. The tradeoff is that on-device AI is small and can read less at once. Apple Intelligence gives about 4,000 tokens of context, while Gemini Nano gives about 12,000 tokens. Because of that limit, Welly is built for focused tasks instead of acting like a general chatbot. It can turn an insurance policy into a plain-language summary, answer follow-up questions about that policy, turn a statement into itemized expenses for review and approval, and help with savings plans, net worth, investments, and a new house affordability calculator.
OpenRouter price checks every few hours showed a sharp rise in GLM-5.2 costs this week. Input rose from about $0.57 to $0.90 per 1 million tokens, and output rose from about $1.80 to $3.08 per 1 million tokens. The change happened through about 10 separate price adjustments over 7 days, with no public changelog or announcement. Tencent’s new Hy3 model also moved around, first dropping and then rising. Low-cost Chinese models can still be much cheaper, as Nex-N2-Mini launched this week at $0.025 for input and $0.10 for output per 1 million tokens. The practical lesson is to avoid locking an AI agent to one hardcoded provider just because it is cheap today. If a model is chosen mainly for price, that price should be watched continuously.
Safari Technology Preview 247 adds the Safari MCP server. It connects an AI coding agent to a Safari browser window, so the agent can inspect how a web page actually appears and behaves. The agent can read the DOM, network requests, screenshots, and console logs, and it can also open tabs, visit URLs, run JavaScript, click, type, scroll, and change the browser viewport size. This helps with Safari-specific layout bugs, slow pages, accessibility issues, form states, and checkout flows without requiring a person to describe every browser detail in a prompt. To use it, Safari Technology Preview must be installed, developer features and remote automation for external agents must be enabled, and safaridriver must be added as an MCP server in a compatible tool such as Claude or Codex. The server runs locally and does not send its own data to Apple. Page content, screenshots, and logs go to the connected agent and model, so it should only be used with tools the developer trusts.
AI agents that can use real tools can get past safety systems that only inspect the request text. A public security vulnerability can be turned into a sequence of tool calls, then rewritten by an LLM as a normal-looking request. The dangerous part may not appear in the words; it appears when the agent follows the tool-call path. In tests with agents using MCP file-system tools, base models from 1B to 14B parameters refused no more than 35% of these attacks. Safety training methods such as DPO and SafeDPO raised refusal to only 48%. Some training-free methods worked better, with one reaching about three times the baseline refusal rate without another fine-tuning run. The methodology, training and evaluation code, dataset, and papers are available.
U.S. companies are using Chinese AI models more often as OpenAI and Anthropic become more expensive to use. Models from DeepSeek and Z.ai are being seen as closer in quality to leading U.S. systems while costing less. On OpenRouter, the share of tokens used by U.S. companies on Chinese AI models has stayed above 30% every week since Feb. 8, 2026, and reached as high as 46%. The average over the previous 12 months was 11%, and it fell to 4.5% in the first half of 2025. Some developers are already using cheaper models for small AI agent projects that need basic decision-making. The broader shift is toward choosing good-enough models by task instead of sending everything to the most expensive model.
NVIDIA Nemotron-Labs-3-Puzzle-75B-A9B is a large language model built for more efficient deployment. It is based on Nemotron-3-Super-120B-A12B and was compressed after training with a method called Iterative Puzzle. The goal is to make interactive, reasoning-heavy, and long-context workloads faster while keeping strong accuracy. The model was reduced from 120.7 billion total parameters to 75.3 billion, and from 12.8 billion active parameters to 9.3 billion. Its design mixes MoE, Mamba, and attention layers, and it supports multi-token prediction for faster text generation. NVIDIA says it delivers about 2 times higher server throughput on one 8xB200 node under matched user-throughput limits. It also raises sustainable 1-million-token concurrency on a single H100 from 1 request to 8 requests. NVIDIA presents it as maintaining strong results on reasoning, coding, multilingual, long-context, and agentic benchmarks.
China may be considering limits on overseas access to its strongest AI models. The concern is whether advanced models from companies such as Alibaba, ByteDance, and Z.ai, including open-weight models, could become harder to use outside China. A counterclaim says recent Ministry of Commerce meetings were mainly about overseas acquisitions and overseas business control, not a broad block on foreign use. The safest reading is that no ban has been confirmed, but Chinese AI models are becoming a more sensitive national and commercial asset. Teams outside China should treat long-term access, licensing, and distribution rules as risks when building on these models.
NVIDIA’s Nemotron-Labs-Audex-30B-A3B is a large language model that handles both audio and text. It is built on Nemotron-Cascade-2-30B-A3B, a text-only base model with 30 billion total parameters and 3 billion active parameters used at a time. The model adds an audio encoder for speech and general audio input, plus audio tokens for speech and other sound output. It can work on audio understanding, speech recognition, speech translation, text-to-speech, audio generation, and speech-to-speech generation. It is designed to keep the base model’s reasoning, knowledge, alignment, long-context handling, and agent abilities with little or no drop. It supports both a thinking mode and a direct instruct mode. It follows the ChatML format and can handle up to a 1 million-token context.
MiniMax-M3-Free is an open-weight 428B MoE multimodal language model. It supports up to a 1M-token context, so it can handle very large amounts of text in one run. It also includes native video understanding. The listed result is 59% on SWE-bench Pro, with a claim that it can rival Claude and GPT-5.5 at about one-sixth of the cost. It can be self-hosted with Ollama or used through an API on OpenRouter. The material also includes guidance for coding agents, agentic workflows, and deployment.
Tencent Hy3 is now available as an open model collection on Hugging Face. The model has 295 billion total parameters, with 21 billion active parameters used during a run. This is the regular release, not the preview version. The license changed from a more restrictive community license to Apache 2.0. The earlier license had limits that blocked use in South Korea, the United Kingdom, and the European Union, while Apache 2.0 is much more practical for commercial use, modification, and redistribution.
Running a full video analysis every time someone asks a new question about the same file wastes work. A better setup treats the video as a reusable knowledge source instead of temporary input. The pipeline first pulls out the transcript, OCR text, scene breaks, and key frames, then stores each observation with its timestamp. It creates embeddings and builds a local index. Later questions use hybrid retrieval to find only the relevant evidence. Only that evidence is sent back to the language model, instead of sending the whole video through the multimodal pipeline again. The main shift is making the model reason from selected evidence rather than re-understand the entire video each time. The approach is packaged as an open-source project called Watch Skill, with access through MCP, a command-line interface, and a REST API.
AI agents cannot safely trust saved memory or retrieved documents just because the source looks authenticated. An open-source agent memory core tested four defenses: scoring records by value, requiring corroboration, letting newer records replace older ones in a fixed way, and giving credit to records that seemed to lead to success. All four failed when the attacker knew how the defenses worked. The weak point was that the record itself supplied the signals used to judge it, so the attacker could write fake value, fake support, convenient timing, or fake success. Provenance and cost are harder for the writer to fake, but provenance only proves where something came from, not whether it is true. False memory can still enter through a real, authenticated session and pass a source check. Related tests point to the same pattern: RAG can quote the right text while pointing to the wrong source about 30% of the time, source and corroboration metadata can be forged, and a cached fact can become stale even though it still fits the context perfectly. The safer rule is to make memory cheap to store but expensive to use for broader influence, by requiring separate anchored evidence before it shapes an answer outside its original scope.
Someone built and open-sourced a 10MB LoRA adapter for Qwen3.5-4B plus a small orchestration layer. For every query it decides whether to answer directly, search the web, or pull from local documents, and it refuses to make up an answer when it can't verify one. It runs locally on Apple Silicon via MLX, with a GGUF build available for llama.cpp and Ollama. The motivating problem: across seven tested models in the 3-9B range, all of them hit a ceiling when trying to state how confident they actually were — they tend to claim high confidence regardless of whether the answer is right. But that confidence information does exist inside the model, in its internal activations. The adapter reads that internal signal directly and uses it to gate whether a tool gets called. In testing, it caught more of its own errors than the base model's tool-calling did, improving a discrimination metric called d′ by 0.46 (95% CI [0.01, 0.89]). Of the cases the gate flagged that the base model missed, 87% turned out to be genuinely wrong answers. There's also a privacy angle: a two-signal version of the system routes personal queries — like asking what a discharge summary said — to a local retriever instead of a web search, cutting down how often private questions leak out to public search.
OpenCode, an AI coding agent tool, runs a process called 'compaction' when a conversation fills up its available context space, collapsing the entire conversation into a single summary. The problem is that this process throws away almost everything valuable that had built up: tool call records, file reads and writes, decisions made, and facts uncovered, leaving only a generic summary that captures roughly 10% of what mattered. One developer had spent months reverse-engineering Claude Code's internals from minified JavaScript files published on npm, carefully mapping out what symbols meant, how functions worked, the main conversation loop, the terminal UI modules, and the permission system. After a single compaction event, the agent lost nearly all of that accumulated understanding, output quality collapsed, and the agent ended up re-reading and re-deriving the same minified code it had already figured out. To fix this, the developer began building a replacement compaction tool called Magic Compact.
LongCat-2.0 is now available on Hugging Face in INT8 and FP8 weight versions. The model is described as having 1.6 trillion total parameters, but only about 48 billion are active for each token it processes. It was trained with large amounts of 1-million-token context data and is positioned for coding, repository-level edits, tool use, and other AI agent work. Reported scores include 70.8 on Terminal-Bench 2.1, 59.5 on SWE-bench Pro, 77.3 on SWE-bench Multilingual, and 73.2 on FORTE, alongside comparisons with major closed models. It can be deployed with SGLang, but the recommended GPU setup uses 16 H20 cards, so this is not a small local model for ordinary hardware. Its chat template includes tool-calling examples and options to turn thinking mode on or off, with thinking mode off presented as a way to improve token efficiency. The weights use the MIT license, while the model card still warns teams to test accuracy, safety, fairness, and language performance before using it in sensitive settings.
Mistral released Leanstral 1.5, a free model under the Apache-2.0 license. It is presented as a 119B-scale model with 6B active parameters, meaning only part of the model is used for each run. Its main strength is formal verification, where software, code rules, or mathematical claims are checked with strict logic. It reached the limit of the miniF2F benchmark, solved 587 of 672 PutnamBench problems, and scored 87% on FATE-H and 34% on FATE-X. Training used mid-training, supervised fine-tuning, and reinforcement learning with CISPO. It is meant for automated theorem proving and formal proof engineering, and in real code checks it found 5 previously unknown bugs across 57 tested repositories.
sim-use is a command-line tool that lets AI agents read and control screens on the iOS Simulator and Android emulators or devices. Instead of sending a large raw screen structure, it turns the screen into a short outline of visible items such as buttons, text fields, and tabs. That outline is about 16 times smaller than the raw accessibility tree, so an AI model can understand a whole screen with fewer tokens. A typical loop is simple: run `sim-use ui` to read the screen, tap an item with a short name such as `tap @9`, then read the screen again to check what changed. It supports several ways to choose what to tap, including fast numbered aliases, stable IDs, visible labels, and coordinates as a backup. After the first command, a per-device background daemon reduces setup work, and each read-and-act round trip is described as taking about 300 milliseconds. The same command style works across iOS and Android, with support for taps, swipes, typing, paste, screenshots, video recording, and app state checks. It can be installed with Homebrew, and it also includes an agent skill that teaches an AI client how to use the tool.
Chimera is an open-source AI agent released under Apache-2.0. For hard tasks, several LLMs answer the same prompt, a judge model compares where they agree or conflict and what they missed, and a final model combines that review into one answer. Easy tasks and tool-use steps stay on a single model, so a cost-and-speed router only spends extra money on multiple models when it is expected to help. The agent loop plans work, acts, checks the result, and rolls back work that fails verification. Verified execution results are treated as the source of truth, so a manager component is not allowed to throw away work that has already been proven correct. Its memory system stores and searches past facts and sessions with SQLite and FTS. It also includes rules for allow, warn, block, and review decisions, plus a static check for cases where the agent tries to modify itself. Chimera can connect to tools through MCP and OpenAPI imports, and it can split work across isolated subagents that each pass their own verification gates. LiteLLM lets it work with OpenAI-style model endpoints, including local models. The project is still alpha: it has 469 tests and passes strict code checks, but it has not yet been proven in production use.
Kimi K2.7 Code is now generally available in GitHub Copilot. It is the first open-weight model that users can choose from the Copilot model picker, and GitHub is positioning it as a lower-cost option for coding work. GitHub hosts the model on Microsoft Azure. Under usage-based billing, the price is $0.95 per 1 million input tokens, $0.19 per 1 million cached input tokens, and $4.00 per 1 million output tokens. Copilot converts this use into GitHub AI Credits, where 1 credit equals $0.01. The rollout starts gradually for Copilot Pro, Pro+, and Max, with support across Visual Studio Code, Visual Studio, Copilot CLI, Copilot cloud agent, Copilot App, github.com, mobile apps, JetBrains, Xcode, and Eclipse. Copilot Business and Enterprise support is coming over the next few weeks, but access is off by default and administrators must enable the policy after checking their security, compliance, and data rules.
Poolside released Laguna XS 2.1, a coding-focused AI model built for agentic coding and longer software tasks. It is a 33B-parameter Mixture-of-Experts model, but it uses about 3B active parameters for each token, which is meant to lower the work needed for each response. Compared with Laguna XS.2, its SWE-bench Multilingual score rose from 57.7% to 63.1%; it also scored 70.9% on SWE-bench Verified, 47.6% on SWE-Bench Pro, and 37.5% on Terminal-Bench 2.0. It trails Qwen3.6 35B-A3B on several listed tests, but Reddit commenters described it as a competitive U.S.-based coding model. Poolside says it can run on a Mac with 36 GB of memory, and it offers quantized FP8, INT4, and NVFP4 versions for tighter hardware budgets. A helper model called DFlash roughly doubled tokens per second in Poolside’s tests. OpenRouter and Poolside’s API serve it with a 256K token context window, and paid pricing is listed at $0.10 per 1M input tokens, $0.20 per 1M output tokens, and $0.05 per 1M cache-read tokens. It supports Ollama, llama.cpp, TensorRT-LLM, vLLM, SGLang, and Transformers, and Poolside also offers a terminal coding agent called pool.
Andon Labs opened a real cafe in Stockholm in April and gave an AI agent control of back-office work. People still made the coffee, while the AI agent handled buying supplies, setting prices, scheduling staff, and talking with suppliers. After two months, the cafe had spent $38,000 and brought in $9,000 in sales. A customer claimed a 99% discount, and the AI agent accepted it without checking. Inventory grew far beyond what was needed. The cafe bought 1,331 pastries but sold only 326, and most of 22.5 kg of canned tomatoes stayed unopened. Press coverage also reported 120 eggs ordered for a kitchen with no stove, plus late-night messages sent to baristas. The practical issue is where human approval should stop the AI agent before it acts.
Claude Code is used as a multi-agent workflow for building a digital product under a 24-hour goal. The process covers research, product choice, planning, building, review, and launch preparation. It uses slash commands, a stop hook, and several subagents with separate jobs. The roles include researchers, draft writers, a red-team reviewer, a judge, skill writers, copywriters, and self-review agents. The example goal is an iOS shipping pipeline product, with only limited human involvement needed near final deployment. The workflow is labeled around quality control, token saving, debugging, shipping, hooks, skills, subagents, and multi-agent work.