Executive summary
An unbranded 'stealth' model that claimed frontier-beating coding performance scored 58.4% on the full 113-task suite versus a cherry-picked 80% headline, according to AINewsOfficial's independent testing, while OpenAI's decision to open-source its Codex execution harness under Apache 2.0 shifted the real competitive battleground from model weights to orchestration layers, per AI Revolution. Stripe's $7.5B acquisition of OpenRouter — 5.8x its valuation from one quarter earlier, according to Nate B. Jones — confirms that routing and metering infrastructure, not any single model, is where durable technical moats are now forming.
LEAD STORY: THE BENCHMARK YOU'RE READING IS PROBABLY LYING TO YOU
An unattributed model calling itself 'Stealth/OX Alpha' appeared on OpenRouter and OpenCode on August 20, offering free inference at a claimed 100 trillion tokens/day — per AI Revolution's reporting, roughly 100x Visa's monthly AI token burn — with zero vendor attribution. Fingerprinting (video tokenizer signatures, a fixed 75-token text offset, emoji rate, censorship-refusal patterns) points with contested confidence to Zhipu AI's unreleased GLM 5.x multimodal line, six days after Zhipu shipped GLM 5.3 as text-only, according to AI Revolution. This is the sixth unclaimed 'stealth' release in six months; the prior five were all later confirmed as Chinese lab launches (Zhipu, Xiaomi, Ant Group, MiniMax) — a repeatable pattern of free public stress-testing on OpenRouter ahead of formal branding. The headline metric — an 80% score on a 10-task DeepSWE subset, versus 52-65% for GPT-5.6 Soul, Claude Fable 5, GLM 5.3, and Grok 4.6 — is a marketing artifact: those same models score 96.0-96.4% on SWE-bench Verified, per AI Revolution's analysis. Independent testing reported by AINewsOfficial ran the full 113-task suite and found a 58.4% resolved rate — the number that should anchor any internal business case. Needle-in-haystack testing from the same source found effective context retrieval succeeding at ~934K tokens against a stated 1M window, failing above ~1.005M tokens. Estimated at 744B total / ~40B active parameters (MoE), the free-tier economics work because low active-parameter cost makes 100T tokens/day commercially survivable as a customer-acquisition play — yet it already drove production adoption via Cursor, Zed, and OpenRouter routing catalogs with zero SLA. Data governance is the variable you actually control: OpenRouter's stealth-model terms permit the provider to use submitted content for training/evaluation, while OpenCode's route states zero data retention, per AINewsOfficial: ```python # Route proprietary code only through the zero-retention endpoint ZERO_RETENTION_ENDPOINT = "https://api.opencode.dev/v1/chat/completions" # NOT the OpenRouter default endpoint, which permits training-data use def query_stealth_model(prompt: str
TOOLING & FRAMEWORKS
A noteworthy development in the tooling space is Composio, a meta-connector layer sitting between AI providers (Claude, Codex, Grok) and 1,000+ business applications, reviewed by Ben AI. Where Claude's native Gmail connector exposes 27 tool calls and cannot send email, Composio's exposes 63 — a 133% increase — and adds multi-account support Claude lacks natively: ```json // Claude Desktop config: connect via Composio MCP { "mcpServers": { "composio": { "command": "npx", "args": ["-y", "@composio/mcp-server", "--apps", "gmail,notion,google-docs"] } } } ``` Free tier: 100,000 tool calls/month, 3 team members; $29/month beyond that, per the source (verify current terms directly). Unused MCP tool definitions load into every chat's context window regardless of use — the creator cites ~10,000 tokens burned by inactive connectors in a single chat — so consolidating 30-50 individual MCP integrations into one routed connection is a direct token-cost redu