AI News Briefing — August 13, 2026

Model releases

• [2026-08-12] Alibaba — The Qwen3.8 weights finally appeared on Hugging Face, and they are not the model people were waiting for. Qwen3.8-2.4T-A95B is 2.4T total with 95B active, but text-only, thinking-mode-only, and 262,144 tokens native rather than the 1M default. Vision, non-thinking mode and the built-in tools stay cloud-side in Max. The licence is a bespoke Qwen3.8-Max License, not Apache 2.0. (official)

  Self-hosting plans built off the Max demos need rechecking against what actually downloads: no vision path, and a context ceiling a quarter of the advertised one.

  For ML / Data Engineers: A pipeline that prototyped against Max hits three separate walls on the downloaded weights — image inputs have nowhere to go, the built-in tools are gone, and prompts tuned on the non-thinking path now run through a thinking-mode-only model whose context tops out at 262K.

• [2026-08-12] xAI — Grok 4.6 is aimed at long-running agents, and Artificial Analysis scored it 61 on its intelligence index, level with GPT-5.6 Sol and behind only Claude Opus 5 and Fable 5. It takes second on GDPval-AA v2 at 1,753 Elo and 65.9% on DeepSWE v1.1. List price is $2/$6 per million tokens — but cross a 200K-token prompt and the whole request bills at $4/$12. (official, source)

  That pricing cliff applies to the entire request, not the overage, so a long-context agent loop can double its bill on one turn that drifts past the line.

  For Engineering Managers / Tech Leads: Compaction and transcript trimming stop being a quality knob and become a billing control at this price shape — a policy that keeps agent context under 200K is worth more per month than the intelligence-index point separating Grok 4.6 from the model above it.

• [2026-08-12] DeepSeek promoted V4 Pro 0813 to its flagship slot with no announcement page at all — the benchmarks went out through a WeChat group and reached the rest of us via Reddit and Hacker News. Rates are unchanged at $0.435/$0.87 per million, cached input at $0.003625, 1M context. Its own table puts Fable 5 ahead by 5.3% across nine agent benchmarks; nobody outside DeepSeek has scored it. (official, source) (unconfirmed)

  A flagship swap you learn about from a screenshot is hard to plan against; should the numbers hold up, the unchanged rate card is the more consequential half, since the capability moved without the price following.

Coding agents

• [2026-08-12] GitHub — Installing an Agent Plugin is now generally available in VS Code, Copilot CLI, the Copilot SDK and the Copilot app, six days after the format was published. A plugin bundles skills/, an mcp.json and a plugin.json manifest, with Copilot-only pieces namespaced under com.github.copilot/ so the rest stays portable. Discovery runs through the Awesome Copilot marketplace, on by default. (official)

  managed-settings.json governs plugin availability, marketplace access and MCP allowlists across all four clients at once — the first time an org can answer “which servers may our agents reach” in a single place.

  For Solution Architects: Keep skills and mcp.json in the portable half of the bundle and namespace only the Copilot-specific defaults under com.github.copilot/ — the same plugin then survives a move to whichever client ships a reader next, which is the bet worth hedging while GitHub is the only one.

• [2026-08-12] Zed — Delta is a separate application, not an editor feature: a shared space where agent transcripts and the diffs they produced sit in one threaded document, with comments anchored to lines that keep moving. It syncs through a new DeltaDB, runs in the browser via WebAssembly, and drives Claude Code and other third-party harnesses. Private beta, first invites out. (official)

  Keeping the transcript next to the diff attacks the review problem agents actually create — a reviewer who can read why a change was made stops reconstructing intent from the patch alone.

• [2026-08-12] JetBrains measured what a missing piece of context costs. Agents kept installing packages into the wrong interpreter and polluting system Python, so PyCharm added a skill exposing get_python_environment and configure_python_interpreter. Across six models and 28 tasks, success went from 68% to 98%; one model went 36% to 96%. The skill only returns information — it never rewrites the command. (official)

  A 36-to-96 jump from handing over a path says the failure was never reasoning, and the read-only design is why: nothing had to guess what the agent meant to run.

Agent frameworks & interop

• [2026-08-12] LangChain — LangSmith BYOC on AWS is generally available across 15 regions: LangChain runs the control plane, while traces, datasets, experiments, prompts and sandbox data stay in your VPC on a private EKS cluster reached over PrivateLink. You provision it with a supplied Terraform module and a scoped IAM role. Enterprise plan only. (official)

  Agent traces are the awkward telemetry: they carry whatever the agent read, so PII and internal API responses land in the observability tier by default rather than by mistake.

AI-assisted SDLC

• [2026-08-12] The LiteLLM poisoning has a size now. CloudSEK put roughly 2,500 organisations and 434,000 files in the attackers’ own logs, harvested when versions 1.82.7 and 1.82.8 sat on PyPI for about 40 minutes on 24 March. The payload took environment variables, SSH keys, cloud credentials, Kubernetes tokens and model API keys. Treat installs through 16:00 UTC that day as suspect. (source)

  A 40-minute window five months ago still reaches you today, because the pipeline that pulled it kept running and the credentials it took were long-lived.

• [2026-08-12] Peking University researchers checked whether coding agents follow the AI-disclosure rules projects now write into CONTRIBUTING.md. Across 106 issues from 49 repositories with such rules, agents largely did not — they never think to look, the way a human contributor does. A contribution can pass every test, break the rule invisibly, and get desk-rejected. (official, source)

  Enforcement rests on a checkbox and reputation, so the fix is nearer the harness than the reviewer: put the disclosure step in the agent’s own PR template.

AI cost tracking & telemetry

• [2026-08-12] AWS documented per-caller cost attribution for Bedrock using Cost and Usage Reports 2.0 with IAM principal data switched on, which fills line_item_iam_principal with the ARN behind each inference call. Athena queries then split spend by principal, role, model and token type — input, output, cache read, cache write — and IAM principal tags roll it up by team or project. CUDOS 5.8+ ships the dashboards. (official)

  Turning on principal data expands every CUR row, and Athena bills $5 per TB scanned, so the attribution you gain has a query bill attached to it.

Practice & craft

• [2026-08-11] Google’s pitch for Go is that language choice is now a review-cost decision. A small syntax surface and one canonical gofmt output narrow what an agent can emit; the compiler and integrated toolchain give it a deterministic signal to self-correct against. The claim underneath is uncomfortable: code that passes its tests can still be the thing that trips the next agent to touch it. (official, source)

  Read as a language argument it is self-serving; read as a checklist — one formatter, one build command, a fast type error — most of it is reachable in whatever you already use.

Research worth reading

• [2026-08-12] Agent skills can make things worse, and the harmful ones look relevant. A differential study across SkillsBench and SWE-Skills-Bench attributed 307 skill-induced failures: 125 functional, 182 efficiency regressions. The largest bucket was Excessive Procedure at 97 cases — 67 of those over-verification, 30 heavyweight implementation pipelines. Prompt length did not explain the slowdowns. (official)

  Relevance is the wrong acceptance test for a skill, which is a problem for every library now growing by whatever seemed useful to whoever wrote it last.

• [2026-08-12] Model rankings flip when you change the token budget. Over 56,476 inferences across four models, three benchmarks and seven budgets from 64 to 4,096 tokens, rankings reversed on every benchmark at p < 0.01, and 3–19% of items got less accurate with more budget. An oracle picking per-item left up to 27.8 points on the table. (official)

  Any leaderboard you are reading was run at some budget, and if yours differs the ordering may not survive the trip — a reason to re-run the two finalists at your own limit before switching.

Watch list

• Agent Plugins outside Copilot. Installation is live across GitHub’s four clients; AWS, Anysphere, Microsoft, OpenAI, Vercel and Google co-signed the spec and none has shipped a reader. The evidence is a changelog from one of them, and whichever lands first sets the working interpretation of the format’s vaguer corners.

  Worth insisting the changelog actually describes reading a plugin.json, not another endorsement — and that it covers skills and MCP servers together, since a reader that handles only half would leave the portability question exactly where it is.

• DeepSeek’s announced price rise, day seven — and whether 0813 gets weights. The rate card still reads $0.435/$0.87 with no figure ever attached to the increase, and V4 Pro 0813 shipped without a model card. April’s V4-Pro and July’s V4-Flash both reached Hugging Face; a repository under the DeepSeek org settles it.

  Seven days of an announced rise with no figure attached leaves batch pipelines planning against a rate that could move any morning, and weights landing would at least give the heaviest of those workloads somewhere else to go.

• An independent score for V4 Pro 0813. Every number in circulation is DeepSeek’s own, routed through a WeChat group. Artificial Analysis or LMArena posting a run is what turns the Fable-5-by-5.3% claim into something to plan against.

  Self-reported tables have landed close before and far off before, so the number worth waiting for is the size of the gap between DeepSeek’s own scores and the first outside run, not merely where it places.

• CoreBreak’s Strands gap, day seven. The managed InvokeHarness path is fixed server-side; _has_tool_use_in_latest_message still short-circuits the event loop in the open-source Python SDK, with no CVE and no patch. Anyone self-hosting Strands should assume the guard stays theirs to write.

  With no CVE attached, nothing in a dependency scanner will tell you this applies to you, so tracking it stays manual — pinning the SDK version somewhere visible is the cheap half of that until a patch appears.