Changelog — September 24, 2026

Jev is a new model class, not a new frontier model. You send a state (a support message, a tool call, a page of DOM) plus typed questions, and it returns the answers in parallel with a probability for each: Choice picks one of your options, Score rates against ordered levels, and Noul gives the probability that a yes/no statement is true. It does not generate strings. That constraint is the product: it cannot invent a category you did not supply, and it cannot return a type error. Cardinality goes up to 255 options, images are not supported yet, and the input context is 32K.

The pricing changes what you can afford to ask. $42 per billion input tokens ($0.042/MTok), and output tokens are free — TypeSafe calls them "too cheap to meter". End-to-end response is 70–500ms. Because every question is answered in one parallel pass, adding questions barely changes latency and costs only the extra input tokens. In a tested run of eight playground requests, evaluation time came back between 92 and 214ms for 4,148 input tokens total — well under a cent. TypeSafe's own homepage claims 193.6x faster and 444.6x cheaper against GPT-5.6 Terra, GPT-6 Astra and Fable 5.1 on workflow evaluations; treat that as the vendor's upper bound, not a typical result.

Why an agent site cares. The decisions a harness makes on every turn — which model should handle this, is this tool call safe, does this belong to support or billing, did the agent actually do what it claimed — are exactly System One work, and today they are usually done by asking a frontier model in English and parsing the reply. LangChain's langchain-typesafe package already ships ModelRouterMiddleware (route to a cheap or capable model by criteria you define) and AutoModeMiddleware (classify a tool call before it runs). We have built out the whole thing: the Jev hub, with setup (SDK, OpenRouter without the waitlist, first call) and use cases (routing, tool gating, triage, agent-trace auditing), plus two tested video breakdowns.

The caveat that matters, from the testing rather than the launch post. "Cannot hallucinate" means it cannot return a value outside your schema. It does not mean the value it returns is right. In AICodeKing's tests, a cafeteria-hours question correctly came back as other when that option existed — and came back as sales at 0.31 confidence when the only options were billing, technical and sales. Always include an other or unknown option and a human path, and do not read 90% confidence as "correct 90% of the time" until you have measured that on your own examples.

Announcement → Affects: /jev/, /tools/cost-calculator/, /security/, /benchmarks/

A flagship that costs less than the one it replaces is unusual enough to lead with. claude-opus-5-5 lists at $4 input / $20 output per MTok with $0.20/MTok cache reads, against Opus 5's $5/$25. Anthropic positions it at Fable 5.1's level on most work at about 40% less to run than Opus 5, with a 1M context and 128K max output. Claude Code v2.1.280 made it the default Opus, so if you pick "Opus" in /model you are already on it.

Three breaking changes before you switch. Thinking cannot be disabled — thinking: {"type": "disabled"} returns 400; omit the field and use effort instead. tool_choice types any and tool return 400; use auto with strict tool use. And computer use needs the computer_toolset_20260801 toolset — the older computer_20251124 now 400s on the API and Google Cloud, though it still works on Bedrock. This is the same shape of break Fable 5.1 shipped on September 1, so a codebase that already migrated for that one has less to do.

The Claude Code releases are mostly for people running a gateway. v2.1.281 adds assume_role on Claude apps gateway Bedrock upstreams (the gateway calls Bedrock as an IAM role via STS, optionally in another AWS account and optionally one session per developer) and guardrail: {id, version} to apply an Amazon Bedrock guardrail to every request through them — set it on all Bedrock upstreams or none. Also: "attribution": false in settings.json hides all commit and PR attribution (keep the object form in files shared with older CLIs, which skip a settings file that holds it), claude plugin validate now reports .mcp.json entries that would be silently dropped at load plus insecure URLs, and /insights estimates how many permission prompts auto mode could have handled for you.

Two auto-mode loops got fixed, and one permission bug is worth knowing. v2.1.280 fixes writes through a symlinked path being judged by their in-tree spelling — acceptEdits, allow rules and auto mode could approve a write that actually landed outside the tree. It also stops auto mode retrying forever when a safety check declines to review an action (now denied once) or returns no answer (now backs off and stops after ten). Add CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH to change the 2,048-character cap on MCP tool descriptions.

Full changelog → Affects: /openclaw/, /openclaw/configuration/, /openclaw/cost-optimisation/, /claude-cowork/pricing/, /tools/cost-calculator/

The GPT-6 range now has a middle and a floor. gpt-6-sol lists at $2 input / $0.20 cached / $10 output and gpt-6-luna at $0.10 / $0.01 / $0.50 per MTok, for prompts up to 272K input tokens. Both are reasoning models that take text and images and return text, through the Responses and Chat Completions APIs.

Luna is the number to notice: at $0.10/MTok in, it is a hundredth of Astra's input price, which is why creators kept routing bulk work to it in this month's videos — walking a finished skill down the model list until quality breaks is now a much longer walk. Both are in our cost calculator.

Changelog → Affects: /chatgpt/pricing/, /chatgpt/api-vs-chat/, /tools/cost-calculator/

The release itself is large — 178 direct commits and 2,614 pull requests from 351 contributors — and continues the pattern 2026.9.5 set: notes published in two formats, one for people and a plain-Markdown changelog for agents and tools. This is the second OpenClaw release since we started tracking its own feed rather than inferring OpenClaw news from the Claude Code feed.

Releases → Affects: /openclaw/, /openclaw/setup/, /openclaw/troubleshooting/

Marketplace plugins can now come straight from a git repository. v7.7.9 accepts a catalog source such as git:github.com/owner/[email protected]#subdir, cloning at the pinned ref, so a plugin can be distributed without publishing to npm. Convenient, and worth the same caution as any unpinned supply chain: pin the tag, read the plugin. The same release adds "Close to the Right" to the Agent Manager tab menu, keeping pinned tabs open.

Automatic session cleanup is back (v7.7.6, Settings → Checkpoints), off by default, with a retention period applied across all projects and Kilo clients on the machine. It protects running sessions and anything recently forked, confirms manual cleanup, and reports processed/deleted/failed counts live. v7.7.7 adopts OpenCode v1.18.14 through v1.18.18 improvements — message ordering, conversation-aware compaction, expanded reasoning-effort support — and retries transient connection drops when approving permissions, which previously left agents stuck mid-approval.

Releases → Affects: /kilocode/, /kilocode/setup/, /kilocode/orchestrator/

A patch tag rolling up ~1,800 merged PRs and 5,071 non-merge commits across 5,169 files (+312,961 / −62,855) since v0.21.3, cut so Docker images and Hermes Cloud pick the window up. Curated notes are deferred to v0.22.0 again — the third consecutive window where the code ships and the explanation doesn't. As with v0.21.2, anyone deciding whether to update is choosing between an unknown and staying behind. Update with hermes update.

Releases → Affects: /hermes/, /hermes/troubleshooting/

The fix most likely to affect a working install: #12237 — OpenClaw 2026.9.1 rejects the legacy auth profile NemoClaw generated, so fresh sandboxes on managed inference.local routes failed their first agent turn, and the Doctor repair it suggested could not run. Host NVIDIA credentials are now cleared before sandbox setup spawns child processes. Related: #10384 stops legacy-credential cleanup deleting entries it never verified as migrated.

Also in the window: connect --probe-only now restarts a stopped sandbox owned by a sibling registry root; nemoclaw config export emits completable service templates for attached Ollama and managed vLLM (with image: null rather than an invented digest); WSL2 ARM64 hosts prove every reported GPU before onboarding uses it; and deferred onboarding now covers LangChain Deep Agents Code, not just Hermes. v0.0.129 docs are staged (#12311), after v0.0.128.

Commits → Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/switching-providers/