A Juejin AI deep dive argues that AI agent spending is structural rather than simply price times usage, framing a fourth discipline — cost-control engineering — alongside harness, context and loop engineering. It opens with a reported Claude Code case in which a user's daily usage jumped from 161 million to 1.306 billion tokens ($41 to $493) because server errors triggered automatic retries that Anthropic customer service said count against billing without compensation. The piece then surveys caching hit rates, tool-definition bloat, Code Mode, context clearing, sub-agent isolation and budget circuit breakers, drawing on figures from Anthropic, Manus and ProjectDiscovery plus third-party harness benchmarks showing 7x to 70x cost differences. It concludes that a shorter correct path, not a cheaper model, is the real target.
雷峰网 AI tested GPT 6.1 Sol and Claude Sonnet 5.5 in Claude Code CLI across 3D escape-room development, policy research and a chemistry explainer. Despite identical OpenRouter token prices, the three tasks cost $38.5 and $10.9 respectively, with speed advantages reversing between tasks. GPT 6.1 Sol spent more time archiving evidence, validating outputs and running browser checks, while Claude Sonnet 5.5 favored shorter paths to usable deliverables.
OpenRouter says the Krea 3 image model is available on its platform at a 50% discount through Nov. 10. Citing Krea, OpenRouter says the model can combine a product shot, a person and a location into one 2K image in as little as nine seconds. The supplied report includes no independent measurements of the model's quality or speed.
OpenRouter says its weekly token throughput rose from 1 trillion to 10 trillion in 50 weeks, then reached 100 trillion in another 28 weeks. The company also says it expects to process 1 quadrillion tokens per week soon, although the figures and forecast are company-reported.
OpenRouter reports that companies and individual users spending at least $100 per month now use a median of six model families, twice the median recorded a year earlier. The platform also published a thread and video about what it calls model neurodiversity and why that trend is accelerating.
AIbase, citing OpenRouter data, reported that Chinese AI models recorded 99.55 trillion token calls from Oct. 26 to Nov. 1, up 6.43% week over week. U.S. models accounted for 20.68 trillion token calls, a 7.97% decline, extending the reported Chinese lead to a 27th consecutive week. Global model calls totaled 194 trillion tokens during the period, and four of the five highest-volume models were Chinese, led by DeepSeek V4.1Flash1017.
A Juejin hands-on report describes Google’s open-source Artemis as an Android agent that operates a phone or emulator from natural-language instructions, with step-by-step reasoning, screenshots and session traces in a web console. The author reports completing a multi-step Sam’s Club China workflow on a Windows-connected Redmi K80 using OpenRouter and a DeepSeek vision model. Artemis also offers Flash and Pro execution modes, MCP integration, multimodal perception and an official claim of 99%+ task completion on AndroidWorld.
OpenRouter published a 233-line Python script that uses the OpenRouter Responses API to upload changed files, test or other context files, and a Git diff to a hosted Linux sandbox. A model can run checks to reproduce suspected bugs and propose V4A patches. The script runs locally, in a Git pre-push hook, or in GitHub Actions, and labels findings as confirmed or unconfirmed, with commands and output supporting confirmed results. In OpenRouter’s own tests of four models on a five-bug change, reviews took one to four minutes and cost $0.015 to $0.045. The Files API, Shell tool, and Apply Patch tool are beta features whose behavior may change; models may also miss bugs, overstate severity, or generate broken patches.
The Memdoor report says it rebuilt the GVS5H multi-agent programming method as a Memdoor workflow using Qwen3.8-27B, with five main roles plus a repair worker coordinating through files. Across five passes on the first 10 LiveCodeBench Hard problems, the workflow averaged 9.8 solved problems, compared with 9.6 for the authors’ harness and 8.0 for single-call Fable 5 results. However, the Fable 5 runs did not use the same harness, and the report says training contamination could not be ruled out.
A Basalt developer reported an experiment with Qwen 3.8 Flash Next that tested two opt-in, lossy inference options: routing bias, which favors experts already on the GPU, and tail skip, which omits lower-ranked experts not resident on the GPU. Both reduce transfers from RAM but can change outputs. In a simulated small-card configuration, the shipped settings increased prose throughput by 56%, while ladder and fact scores fell by 5.2 and 2.3 points, respectively. The developer’s deliberately worst settings reduced fact accuracy from 38.1% to 2.7%. The test covered one model on one rig, used one seed for ladders and three for facts, and measured speed with single cold runs. Basalt ships both flags disabled by default and warns when they are enabled.
Darkbloom told members of its decentralized AI inference network that falling token prices could reduce revenue from spare Mac compute capacity. The network estimated that an M5 Pro Mac Mini with 64GB of unified memory could earn up to $192 monthly under the best-case scenario, but about $36.77 after accounting for a reported 77% decline in Anthropic’s price per million tokens. Darkbloom is analyzing model demand, exploring workloads beyond inference, improving network performance, and seeking clients outside OpenRouter.
Cisco Talos researchers say VOLTA-AUTODOXER is a Windows infostealer that steals credentials, sessions, payment data, cryptocurrency information and other sensitive material before building a victim dossier. It sends the serialized dossier through OpenRouter, defaulting to Claude Opus 4.6 and falling back to Kimi K2 and Gemini 2.5 Flash if a model refuses or returns inadequate output. Without an OpenRouter API key, it can skip the AI assessment and exfiltrate the raw dossier. The malware's main executable may delete itself after infection. Talos did not disclose where the malware was found, how widely it is distributed or how many devices it infected.
nullprogram introduced Prospero, an open-source C++ coding harness with a Dear ImGui interface, desktop control, subagents, remote deployment, and operation deadlines. The Community Edition is free, while the Professional Edition adds embedded headless engines for macOS, Windows, and Linux; users must supply a model provider. The author claims Prospero uses around two orders of magnitude fewer resources than competing harnesses, but no independent benchmark is provided.
OpenRouter reports that open-weight models rose from 27% to 72% of the top 50 models’ tokens on its platform between October 2025 and late September 2026. Its token-weighted pricing data lists open-weight models at $0.28 per million input tokens and $1.62 per million output tokens, compared with $1.27 and $6.36 for closed-weight models. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scored 57.6, compared with 46.3 for MiMo-V2.6-Pro, the leading open-weight model in the comparison. The report says these scores depend on reasoning-effort settings, the 11-point gap is a snapshot, and token counts use providers’ own tokenizers. OpenRouter recommends testing an open-weight model on a team’s tasks and using closed-weight models for steps where tests show extra quality matters.
San Francisco startup The Biological Computing Co. released TBC Mischief 1, an image-to-video model built on Alibaba's open video model Wan 2.2. It turns a still image and prompt into a 3- or 5-second clip at 480p or 720p. The model is available on fal and OpenRouter for $0.005 per second of generated video, so a five-second clip costs 2.5 cents. TBC says insights from lab-grown living neurons make the model about five times faster than Wan 2.2 at the same quality, but it has not published independent tests supporting the claim.
A V2EX user reported that a Google account subscribed to Claude Max since January was banned in mid-October. The user then said another Gmail account was banned after identity verification with a Chinese passport, while a newer Gmail account was banned immediately after a successful payment. All three attempts reportedly used the same device, US credit card, and network proxy, but the cause of the suspensions remains unconfirmed.
OpenRouter gave Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 the same 11 hidden-test coding tasks and a Claude Code bug-fix task. All three passed every test, while Haiku cost 19 to 30 times less per task than Opus and Sonnet cost about half to two-thirds as much. OpenRouter said the small, well-specified tasks may not reflect long-running, ambiguous agentic work, and recommended testing models on each team’s own workflows.
OpenRouter published session-cost data for Claude Code from its own platform telemetry covering the 30 days to October 18, 2026, reporting median costs of $0.31 for a 2-to-9-turn session on Claude Opus 5.5 and $0.16 on Claude Sonnet 5.5, with under a cent on Haiku 5.5. In a controlled bug-fixing test on a small invoicing library, all three Claude models fixed every bug, with median per-session costs of $0.098 on Opus 5.5, $0.062 on Sonnet 5.5 and $0.0052 on Haiku 5.5. The post also shows how to run Claude Code through OpenRouter's Ori Harness with per-session model selection, a cost-tracking statusline and credit limits or workspace budgets.
OpenRouter shipped its largest provider-router update since its founding, replacing a system that mostly prioritized price with one that rewards inference quality at each request's actual cost. OpenRouter says the new architecture considers cost, throughput and accuracy, and was developed through months of research and testing. Provider Eitan Borgnia said the update lets providers differentiate across those axes and encouraged improvements to serving engines.
OpenRouter reports that Riverflow 2.6 Pro is now available on its platform. The vendor describes it as a major upgrade to Riverflow 2.5 Pro, focused on real-world capabilities for creative teams and priced at the same rate.
OpenRouter and ElevenLabs announced a livestream billed as an “LLM Battle Royale.” The supplied report does not provide details about participating models, technical findings, competition results or a winner.
Stu Cao said Prime Inference scaled from zero to serving 120 billion tokens per day within one month of being onboarded to OpenRouter. He added that this represents only a small portion of Prime Inference’s overall activity. The figure comes from a reposted statement and has not been independently verified; the report provides no measurement methodology or underlying usage records.
OpenRouter announced the return of “Royale: Last Agent Standing,” a livestream in which AI models from multiple labs compete across 30 games, with one winner. The post scheduled the event for October 29 at 12 PM ET and said live casting would be powered by ElevenLabs V4.
OpenRouter evaluated 25 prompt-injection attacks against four models using a document-summarizing assistant, comparing configurations with defenses enabled and disabled. Its Armor tool-call filter reduced successful attack emails from 73 of 1,440 attempts to zero, while labeling documents as data alone left 13 of 720 attempts successful. Armor allowed nearly every explicitly requested email but blocked one legitimate request because the user had not named the recipient address.
OpenRouter published a guide on October 29, 2026, explaining how to send PDFs to file-reading LLMs with a strict JSON schema that produces typed table rows. In its first-party test of five PDFs, including two scans, every model that read the PDFs directly extracted all 644 cells correctly; the guide also covers OCR routing, empty results, output-token limits and cross-page tables.
Particle-1, a token-efficient coding model post-trained on DeepSeek V4.1 Flash, is now available on OpenRouter. Particle reports that it uses about 45% fewer tokens on average while maintaining accuracy close to the base model.
Kenny Rogers ran the same single-file HTML solar-system prompt across four models on OpenRouter to compare speed, cost, and output. Gemini 4 Argon reportedly finished fastest in 42 seconds at a cost of 3.3¢, while DeepSeek V4.1 Flash was the cheapest at 0.6¢ and took 85 seconds.
OpenRouter says adding the X-OpenRouter-Cache: true header saves the full response and serves an identical repeat request in about a second at no charge. According to the vendor, this skips the model entirely, unlike prompt caching, which still runs it. The cache lasts five minutes by default and can be extended to 24 hours with X-OpenRouter-Cache-TTL.
OpenRouter says users will not be charged for failed generations, including text requests with zero output tokens and a blank finish reason. The company also says error-terminated generations and broken or empty image outputs consume zero credits, with the policy enabled by default for all models even when upstream providers still charge OpenRouter.
OpenRouter data cited by Goldman Sachs’ Bobby Molavi shows open-weight models handled 72% of tokens in the second week of September but received only 27% of platform spending. The report also says token demand rose sharply while GPU rental prices increased and token prices fell, creating a disputed divide between infrastructure revenue and model monetization. It further cites roughly 69% of mid-September tokens as coming from Chinese AI laboratories and reports that Meta sought up to $95 billion from shareholders for AI capital expenditure.
What problem it solves
AI agents need to choose among models while balancing task success, cost, and latency for their workloads. Applications also need flexibility in where routing runs and who controls model clients, credentials, retries, and service operation.
Key features
Multiple routing algorithms — Routing can be configured with Auto, Task, Execution, or Composite strategies, including mode…
What problem it solves
Researchers, red-teamers, and system tinkerers need a chat interface that can work with multiple hosted and local models while exposing evaluation, perturbation, and sampling controls. They also need clear visibility into data flows, storage, telemetry, and deployment boundaries so they can make informed choices about privacy and self-hosting.
Key features
Multi-model races …
What problem it solves
Useful software engineering agents can require complex control flows, specialized tools, extensive configuration, and large dependency footprints. This project provides a simple, hackable agent that can work through the shell and run in local or sandboxed environments.
Key features
Shell-only agent — The agent interacts with its environment through bash without requiring mod…
What problem it solves
Teams that rely on AI agents need a way to give each agent a role, workspace, and instructions, keep context across conversations, and control what agents are allowed to do. Without that, agents run blind: actions happen without review, memory does not persist, and repeating scheduled work means redoing setup by hand.
Key features
Create agents and teams — Give an agent a na…