compact_before_next_turn() in the tool runner, wrapping the September 14 Messages API beta into a single ergonomic method call — the full compaction story from API to SDK to tool runner is now complete. Second, E2B overhauled its SDK defaults: e2b@2.51.0 removes preset timeouts, fork counts, and secure flags in favor of v2 API defaults, and the desktop SDK's allow_internet_access flips from True to not set — any script that relied on those defaults changes behavior on update. Separately, Qwen3.8-Omni-Flash lands as a serious multimodal-agent platform (1M context, omnimodal, 98% cheaper audio, tool use, API-only) and two new arXiv papers independently confirm that frontier coding agents regularly fail to do what they report doing — the overclaiming propensity paper puts a 67.9% file-skip rate and 80.4% misleading-response rate on the problem that every developer relying on agent output should know. And Bend 2 launched September 17 to 589 HN points: a language that makes AI-generated code violations of declared invariants fail at compile time.allow_internet_access flips — check any script relying on default 5-minute timeout or desktop internet access before updatingcompact_before_next_turn() — the ergonomic wrapper for the on-demand compaction beta is now in both major SDKscurl -fsSL https://bend-lang.com/install.sh | sh
cat > LAWS.bend << 'EOF'
law no_negative_balance (acc : Account) :
acc.balance >= 0
EOF
bend build
allow_internet_access previously defaulted to True for desktop sandboxes. It now passes no default, meaning the API's own default applies — which is no internet access. Any desktop sandbox created without explicitly passing allow_internet_access=True will now be isolated from the internet on update. The secure option is also deprecated (still accepted but ignored) — all sandboxes use v2 API with secured envd access by default.allow_internet_access: True default — desktop sandboxes now create without internet access unless you explicitly opt in; secure parameter is deprecated (ignored).allow_internet_access defaulting to True (for browser automation, API calls, or anything requiring outbound internet from the desktop sandbox), add allow_internet_access=True explicitly in Python or allow_internet_access: true in TypeScript. (2) Remove secure parameter calls — the parameter is deprecated and silently ignored. All sandboxes now use secured envd access via the v2 API endpoint. (3) Also check main SDK usage (e2b@2.51.0) — the 5-minute sandbox timeout default is gone from the client; pass an explicit timeout parameter if your code relied on the client-side safety cap. (4) Client-side fork count validation removed — the API now validates fork count directly. Edge-case fork operations at the old client-side validation boundary may now surface as API errors rather than client-side rejections.secure parameter on sandbox creation.EffortModerate — add allow_internet_access=True explicitly to desktop sandbox creation where needed; add explicit timeout to main SDK sandbox creation if you relied on the default; remove secure parameter.qwen3.8-omni-flash via the Qwen API. If you need open weights, Qwen/Qwen3.8-Flash-Next is on Hugging Face (the base architecture without the full omnimodal stack). (5) Tool use is built in — not an afterthought; this positions Qwen3.8-Omni-Flash as a direct competitor to Gemini 3.8 Flash for multimodal agentic pipelines.CLAUDE_CODE_AUTO_MODE_SERVER=0. A new Auto mode server row appears in /status showing whether the current session's classifier runs on the server./status row showing classifier location, and adds CLAUDE_CODE_AUTO_MODE_SERVER=0 opt-out for cloud providers./status after updating to verify Auto mode server: true. (2) Bedrock/Vertex/Foundry/gateway users: if you have strong reasons to keep the classifier client-side (e.g., latency profiling shows client-side is faster for your workload), set CLAUDE_CODE_AUTO_MODE_SERVER=0 in your environment. (3) Billing transparency: The auto mode classifier billing details and server-side default are now documented at code.claude.com/docs/en/auto-mode-classifier-billing — if you have cost-sensitive agent pipelines, read that page. (4) Most users: no action required — this is a default change that benefits you; only act if you specifically need client-side behavior.npm install -g @anthropic-ai/claude-code@latest; confirm /status shows the new row; optionally set CLAUDE_CODE_AUTO_MODE_SERVER=0 to opt out.compact_before_next_turn() / compactBeforeNextTurn() to the tool runner — a single method call that triggers on-demand compaction before the next model turn in an agentic loop. The rate limits API adds group.display_name to replace group_type (deprecated, not yet removed). TypeScript additionally fixes BetaMessageStream incorrectly coercing a null compaction_delta to the string "null", adds TypeScript 6.0.3 build internally, and switches unit tests from jest to vitest. Python fixes multiple X | Y union type parsing edge cases and joins multiple anthropic-beta values with comma+no-space correctly.compact_before_next_turn() as the ergonomic on-ramp to on-demand compaction in agentic tool loops, and deprecate group_type in favor of group.display_name in rate limits.compact_before_next_turn() wraps the compact-2026-09-04 Messages API compaction beta into a single method. Call it inside your tool runner loop at the point where you expect context to grow large — before the expensive turn, not after. The runner handles building the compaction request and wiring the signed block. This is the intended ergonomic path for the on-demand compaction feature. (2) Rate limit monitoring migration: If you parse group_type from rate limit response headers or objects, switch to group.display_name — group_type is deprecated and will be removed in a future release. The display_name value is a human-readable string identifying the rate limit group. (3) TypeScript BetaMessageStream null fix: If you were matching on the string "null" in a compaction_delta field from a streaming response, update your parsing — null values are now returned as actual null, not the string. (4) TypeScript builds with TypeScript 6.0.3 internally — no change required in your code; note only if you have strict peer-dependency constraints.group_type.EffortQuick — pip install anthropic --upgrade / npm install @anthropic-ai/sdk@latest; add compact_before_next_turn() calls at appropriate loop points; migrate group_type to group.display_name in rate limit monitoring code.TypeSafeModel now fills tool arguments when Jev can express them (Jev can now drive tool-use parameterization directly, not just text responses); union output types are handled by first selecting the type; new typesafe_boolean_threshold parameter controls what yes/no confidence level registers as True. New Choices helper builds option sets at runtime for Jev decisions. TemporalDurability gains event_stream_topic for Temporal Workflow Streams. RealtimeSession.wait_for_playback() added. UseEnumMemberDocstrings for enum member docstring descriptions. Fixes: ToolReturnPart recording during realtime tool errors; adaptive thinking with tool output on Bedrock; gateway ID misclassification as OpenAI o-series models; invalid URLs from web_fetch_tool now trigger ModelRetry.typesafe_boolean_threshold: Set the confidence level above which Jev's yes/no response registers as boolean True. Useful for high-stakes boolean decisions where you want Jev to be more confident before committing. (3) Choices helper: Build a list of valid options at runtime (from DB, config, live state) for Jev to select from — replaces static enums when the option set is data-driven. (4) Temporal Workflow Streams: Add event_stream_topic to your TemporalDurability config to stream agent events during execution via Temporal's Workflow Streams feature — connects pydantic-ai durable agents to Temporal's streaming infrastructure. (5) Bedrock adaptive thinking fix: If you've been running Bedrock agents with thinking_tokens and tool use and saw failures on tool responses, this release fixes it. Update and re-test Bedrock flows before next prod deployment.pip install pydantic-ai --upgrade; add TypeSafeModel tool-argument filling if using Jev; tune typesafe_boolean_threshold for boolean tool arguments; add event_stream_topic if using Temporal Workflow Streams.Sandbox.create() no longer presets a 5-minute timeout; Sandbox.fork() no longer presets count: 1. Sandbox operations now use v2 API endpoints with secure envd access by default. Client-side fork count validation removed — the API validates directly.count: 1 preset, moving fully to v2 API endpoints with secure envd; explicit timeout parameters are now required if you relied on the 5-minute default.timeout parameters if your sandbox creation relied on the implicit 5-minute timeout as a safety cap. Without it, sandbox lifetime is now governed by your API account defaults. Example: Sandbox.create(timeout=300) for a 5-minute cap. (2) Fork operations: If you called Sandbox.fork() without count and relied on count: 1, this now passes no preset. Verify your fork workflows handle the API's default (or pass count=1 explicitly). (3) v2 API migration complete: All sandbox operations now use v2 API endpoints. If you have monitoring or code that pattern-matches on endpoint URLs, update it accordingly. (4) Note: The secure parameter on Sandbox.create() is also deprecated in the main SDK (consistent with the desktop SDK — see Breaking Changes).Sandbox.create() without an explicit timeout; you call Sandbox.fork() without an explicit count; you monitor E2B API endpoint patterns.EffortModerate — add explicit timeout to sandbox creation calls where you need a time cap; verify fork behavior without the count: 1 preset.LAWS.bend file where you declare formal rules the program must always satisfy. After every code change (including AI-generated changes), the Bend compiler re-proves these laws as theorems. If any declared invariant is violated, the compiler refuses to produce a binary. Proof checking runs in under a second on properties that take other tools minutes. Compiles to C, Metal, CUDA, and JavaScript. Apache-2.0 license.compact_before_next_turn() method now in the Anthropic SDK (today's API & SDK Changes). Install as a Claude Code plugin; the marketplace listing provides a verification layer. Validate the repo ownership and star-history momentum before committing — the star count alone does not confirm sustained growth vs. a launch spike.The SWE-bench Verified leaderboard expanded from 78 models (September 18 snapshot) to ~116 models between scans, with no change in top rankings. Claude Opus 5 leads at 96.0–97.0%, GPT-5.6 Sol at 96.2%, Claude Fable 5 at 95%. The top 5 span less than 5 percentage points — SWE-bench Verified is saturated. SWE-bench Pro (74 models) remains the more informative ranking: Claude Fable 5.1 leads at 81.2%, Claude Mythos 5 at 80.3%.
No new model has entered the top 3 of either leaderboard in the September 18–19 window. No new Chatbot Arena ELO ranking movement confirmed in the window.
201 points and 57 comments on the harness design ablation paper. Practitioner discussion confirmed the bash-only finding: multiple commenters reported the same cost advantage in their own setups when dropping predefined-tool scaffolds for bash-capable models. The most-upvoted thread: "the predefined tool interface is a crutch for models that don't know bash, not a signal of a good agent harness." Notable dissent: one commenter argues predefined tools provide human-reviewability and auditability that a bash-only transcript lacks — a valid concern for regulated environments. The discussion validates that the paper's findings map to production experience, not just benchmark artifacts.
589 points and 302 comments — the highest-scoring developer-tool post of the day. The thread is the most comprehensive community discussion of where LAWS.bend formal invariants can and cannot reach. Key signal threads: (1) users with dependent-type experience confirming proof checker quality; (2) discussion of what class of business-logic invariants Bend can and cannot express (spatial logic, temporal properties, and cross-contract invariants are harder); (3) skepticism about "runs on GPUs" claim being for computation, not verification. Victor Taelin responded to claims that Bend is "vibe-coded formal verification without domain knowledge" — confirmed 10 years of full-time dependent-type research.
APIError for mid-stream exception/error frames on Bedrock; switches to TypeScript 6.0.3 and vitest internally. [github.com/anthropics/anthropic-sdk-typescript]Migrate to antigravity-preview-09-2026, rename all snake_case built-in tool parameters to PascalCase, and update file-edit handling from full-rewrite to line-range model. See the September 18 digest Breaking Changes for the full migration guide. [ai.google.dev/gemini-api/docs/changelog]
Standard pricing: $10/$1 (cached)/$50 per MTok. Long context (>272K tokens): $20/$75 per request. Free billing ends October 4. Confirm billing account and cost model before that date if you're running GPT-6 Astra in any pipeline. [platform.openai.com/docs/changelog]
Jev (TypeSafe AI) is currently waitlisted at typesafe.ai; no GA date given. pydantic-ai integration (v2.45.0, v2.46.0) is available now for existing waitlist users. [typesafe.ai]
V4.1 Pro confirmed in development as V4 Pro successor. [api-docs.deepseek.com/updates/]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.