OpenRouter joins Stripe, Ornith-1.5 hits coding benches, AWS wires Step Functions to agents (2026)

Morning scan for people who ship with coding agents. Window covered: roughly the last 36 hours ending early Aug 20, 2026 (Asia/Shanghai).

Overnight

OpenRouter is joining Stripe. Same product for now; deal still needs to close.1
Ornith-1.5 landed as an open model family aimed hard at coding and agentic work, with a 397B flagship and a 9B phone-sized cut.2
AWS put a Copy agent prompt button in the Step Functions console so Claude Code, Cursor, Copilot, Codex, and friends can self-wire ASL skills plus an MCP server.3
OneCLI launched on HN as an open-source, sandboxed agent harness built for whole teams, not solo demos.4

Headlines

OpenRouter is joining Stripe

OpenRouter says it will keep the same name, product, and roadmap after joining Stripe, and that routing stays driven by what’s best for the user rather than any single model lab.1 The company positions itself as a multi-model gateway and marketplace — one interface across 400+ models, with observability, cost controls, and routing — and claims 10+ trillion tokens per day for a community of over 10 million developers and companies.1

If you already point agents at OpenRouter, nothing in the post asks you to change SDKs or keys today. The interesting bit for coding stacks is the parent: Stripe’s fraud, billing, and enterprise distribution muscle sits next to a neutral model router. Whether that neutrality holds after the close is the thing to watch, not the press-release promise that “nothing changes.”

The transaction is still subject to customary closing conditions; OpenRouter expects to close in the coming weeks.1

Ornith-1.5: open weights that chase Claude-class coding benches

Ornith’s 1.5 release extends its self-scaffolding idea into a fuller self-improvement loop: the model proposes new tasks, builds task-specific scaffolds, and produces solution rollouts for RL.2 Three sizes shipped: 397B MoE, 35B MoE, and 9B dense (plus a mobile quant of the 9B).2

On the numbers that matter if you care about agentic coding:

Model	Terminal-Bench 2.1	DeepSWE	SWE-bench Verified
Ornith-1.5-397B	86.1 (Terminus-2)	56.0	86
Claude Opus 4.8	85.0	59.0	85.8
DeepSeek-V4-Flash-0731	82.7	54.4	81.6
GLM-5.2	82.7 / 81*	46.2	83

*GLM-5.2’s Terminal-Bench figure is 82.7 in the headline comparison and 81 under Terminus-2 in the full table; both come from Ornith’s post.2

The smaller cuts are the sleeper story. Ornith-1.5-35B (3B active) posts 68.5 on Terminal-Bench 2.1 with Claude Code and 79 on SWE-bench Verified; the 9B hits 47.0 and 70.6 on those same two.2 If the eval harness notes hold up outside their paper, that’s a serious open alternative for local and mid-tier agent loops — not just another chat benchmark dump.

AWS turns Step Functions into an agent-native workflow surface

AWS added a Copy agent prompt control in three places in the Step Functions console (home, Create State Machine modal, Local Development).3 Paste it into your agent and it installs two pieces: serverless skills from the Agent Toolkit for AWS (ASL, retries, saga/fan-out patterns, SAM/CDK deploy guidance) and the AWS Serverless MCP server so the agent can create state machines, start executions, and inspect history.3

Supported wiring out of the box includes Claude Code, Kiro CLI, Codex, Cursor, GitHub Copilot, Devin Desktop, and OpenCode, plus a generic MCP JSON block for anything else.3 Credentials stay on your existing local AWS profile.3

This is the kind of boring infrastructure win that compounds: less time teaching the agent ASL from scratch, more time describing the workflow you actually want. Available in commercial AWS Regions at no extra feature fee.3

OneCLI: one sandboxed agent per teammate

OneCLI (YC S26) is launching as an open-source platform that gives every employee a personal agent in an isolated sandbox, with credentials injected at a gateway so the model never holds the real secret.4 5 Org policy is enforced outside the model at the network layer; approvals for high-risk actions (send email, delete a Linear ticket, empty a bucket) sit in the chat itself.5

The stack is self-hostable: web dashboard, API control plane, Rust gateway, runner, sandbox supervisor, optional Slack channel adapter.5 License is Apache-2.0 with enterprise-only paths under a separate eval/production license.5 If your company has been gluing OpenClaw-class agents together with shared API keys in plain text, this is aimed squarely at that mess.

Also shipping

Frugal Tokens maps where coding-agent spend actually goes

Show HN: Frugal Tokens inspects coding-agent sessions for cost, cache reuse, overlapping work, and cache-miss causes (TTL, compaction, model changes, unexpected full/partial misses).6 The public demo surfaces session-level breakdowns — model mix, turns, processed input, reuse rate — so you can see which sessions ate the budget instead of guessing from a monthly invoice.7 Repo: github.com/dpclark4/frugal-tokens.

Cerebras CS-4 pushes interactive inference further

Cerebras introduced CS-4, a rack-scale system it claims delivers up to 30× faster inference than GPU systems and more than 1,000 tokens/sec on models over 10T parameters, with first shipments starting this quarter.8 For coding agents, raw decode speed is not a feature list item — it is how snappy multi-tool loops feel when context is huge. Treat the 30× claim as vendor-stated; the page itself notes results vary by workload and configuration.8

Repo watch

mattpocock/skills — “Skills for Real Engineers.” Small, composable agent skills (/grill-me, /tdd, /to-spec, architecture surveys) meant to fix common coding-agent failure modes instead of handing the whole process to a mega-framework. Install via Claude Code plugin or npx skills@latest add mattpocock/skills.9 It was near the top of GitHub’s daily trending list as this issue was built, with heavy star velocity in the daily window.10

Read if you have 10 minutes

Jeremy Morrell’s Extensible software in the age of LLMs argues that once users can speak code into existence, products should ship a solid core plus safe extension points (capability-based sandboxes, not raw API keys in the prompt).11 Pi, DeepSeek Harness, and OpenCode’s “everything is a plugin” layout show up as local-first examples; the harder problem is doing the same on the multi-tenant web. Worth it if you are designing agent harnesses or internal platforms rather than just consuming them.