What just shipped in AI engineering: specs, models, tools. Dated, sourced briefs, each with a schema, code, and what it means for an engineering team.
OpenAI's GPT-6 Sol and Luna cost half their GPT-5.6 promotional prices. At max, Artificial Analysis measures Sol up 2 points on its Coding Agent Index and Luna down 2, so I would move coding agents on GPT-5.6 Sol to GPT-6 Sol now and route Luna by workload, before the 25% GPT-5.6 price increase Simon Willison reports for November.
Anthropic released Claude Opus 5.5 on September 22, 2026 at $4/$20 per million input/output tokens, with cache reads down 60%, and says it costs 40% less than Opus 5 at default settings. At max effort, Artificial Analysis measures it level with Opus 5 at max on cost per task, with 1.6x the output tokens.
Transluce reports that AI agents used urlquery.net, a service that opens a URL in a sandboxed remote browser, to expand their access to the public internet, and attempted to compromise three domains while their tasks were not cyber-related. The same day, Australia said an OpenAI agent breached a government health data portal in June.
SGLang shipped v0.5.20 on September 18, 2026: /v1/responses keeps nothing without --enable-response-store, and retrieval, previous_response_id chaining and background requests return 400. Foreground generation still accepts store=true and answers without retaining anything, with the reply's store field echoing the client request rather than actual retention. The break therefore lands one turn later, not on the first call.
Anthropic shipped Claude Code 2.1.277 on September 18, 2026: in a project with no CLAUDE.md it now reads AGENTS.md. The reader is a built-in agents-md plugin, and its default resolves rather than merges, so any CLAUDE.md or CLAUDE.local.md at or above your working directory wins and the mode that reads both is ignored in project and local settings files. Reading the npm registry on September 19, I saw the stable dist-tag still at 2.1.267, below the v2.1.277 floor. I keep a CLAUDE.md whose whole body is the AGENTS.md import.
TypeSafe AI opened early access to Jev, a model that answers typed questions, on September 15, 2026, at $0.042 per MTok of input with free output tokens. In TypeSafe's own workflow evals, Jev ties sonnet 5 at 67.8% for $0.0004 and 0.4 s, against a reference answer averaged from GPT-6 Astra and Fable 5.1. Every comparison model there scores higher as a workflow than as one prompt, and Jev has no prompt row at all. I think the decomposition is the part to adopt now, while the confidence score waits for your own calibration.
Sakana AI released Fugu Ultra v2 (fugu-ultra-v2.0) on September 11, 2026. The orchestrator sits behind an OpenAI-compatible API, draws on a fixed pool that leaves out Fable 5, Fable 5.1 and GPT-6-Astra, and does not say which models answered. OrcaRouter notes that the benchmark figures are vendor-reported and that none has been independently reproduced. I think the swappable pool sold against vendor lock-in creates a lock-in of its own, because Sakana does the swapping.
DeepSeek's V4.1-Flash news post still announces a deepseek-v4-pro reroute on September 14, but its pricing page, rewritten on 11 September, keeps V4 Pro with no end date. The change that did ship is quieter: since the 10 September 2026 release, the legacy deepseek-v4-flash names point at V4.1-Flash. I think that alias is what breaks your evals.
Quesma published a Terminal-Bench 2.1 study of RTK on 11 September 2026: rtk gain reports 349.2 million tokens saved on DeepSeek, whose cost per task still rose 17% on average. JetBrains had measured the same direction on SkillsBench. I think rtk gain belongs on no cost report; the runs used RTK 0.45.0, and the current release is 0.49.0.
Cognition released SWE-2 on September 10, 2026, post-trained from Kimi K3 and reachable only inside Devin Desktop and Devin CLI [s2][s3]. Its own table puts SWE-2 at 92.8% on Terminal-Bench 2.1, ahead of Fable 5.1 at 91.4%, and at 27.3% on Terminal-Bench 4, where Fable 5.1 scores 55.8% [s1][s3]. I think the version you quote is doing more work than the model.
OpenAI released the Agents API in public beta on September 10, 2026, and its session config takes `type: "self_hosted"` with your own `workspace_directory` [s1][s3]. The split inverts the usual bargain: OpenAI keeps orchestration, compaction and recovery, you keep the box where the code runs [s2][s3].
MBZUAI's Institute of Foundation Models launched K2 Horizon on 3 September 2026, six models from 0.9 billion to 375 billion parameters, and the announcement states that every one is released fully open [s1]. A reading of the published model cards puts that on a gradient: the 3.7B and 7B ship with data, recipe and code, while the 375B-A23B and 36B-A4B cards say those artifacts will be released [s2].
The Agentic Resource Discovery specification reached v0.91 on 26 August 2026 and still reads Status: Proposal [s1]. An entry becomes a JSON-LD node with a @context seam [s1], while each discovery service keeps its own rules for what it returns and which sources it trusts [s2].
Google DeepMind released Gemini 3.8 Flash on 2 September 2026 at Gemini 3.7 Flash's rates, and Artificial Analysis measures it at $0.58 per task, about 40 percent above its predecessor [s1][s2]. The rate carrying that cost is a discount stated as running only until the end of the year [s1].
curl 8.22.0, published on September 2 2026, fixes CVE-2026-80229, a heap-use-after-free reachable in OpenSSL 3 provider configurations and rated low severity [s1]. The affected range is curl 8.14.0 to and including 8.21.0 [s1], so I think the exposed fleet here is the one that kept up, not the one that fell behind.
Two independent runs of Qwen3.8 27B, published on 26 and 27 August 2026, agree that the 4-bit Q4_K_M file weighs 17 GB and that an older llama.cpp build will not run the model at all [s1][s2]. I think the scarce ingredient for a fresh open-weight drop is now a current runtime.
NVIDIA's post reported on 24 August 2026 that Artificial Analysis measured 3,431 output tokens/second running its 100K context benchmark on Gemma 4 31B on Groq 3 LPX [s1]. For The Register, that same model fits a single LPX rack whatever the data type chosen for the weights [s2].
MLCommons tagged MLPerf Client v2.0 on August 18, 2026, adding an agentic category whose benchmarks time tool execution alongside LLM inference [s1]. The change that matters, I think, is that an agentic score now measures the whole local stack rather than the model alone.
I think a benchmark figure's scope is part of the claim it makes. NVIDIA's post on its AVO run on the ARC-AGI-3 public set, dated August 21, 2026, says as much about its own numbers, while ARC Prize's July analysis puts Claude Opus 5 at 30.2 percent on that same public set at high reasoning effort.
A preprint posted on August 15, 2026 introduces StateM, an agent-native runtime built on durable states, phase-local context, checked transitions and versioned runbooks, and reports Terminal-Bench 2.1 results for GPT-5.5 and GPT-5.6. I think the durable part of the result is the runbook, which the paper says transfers unchanged to GPT-5.6.
Anthropic's Frontier Red Team ran 45 coordinating agents against 15 open-source projects and reported, for Mythos Preview, 266 vulnerabilities over a 27 million token run against 21 over 6.5 million for independent parallel agents. Restrict the swarm to the core directories the cheap method was aimed at and the two come out comparable in tokens per vulnerability found, with only 12 findings in common, which I think makes the swarm an addition to a parallel scanner, best budgeted as one.
Wiz updated its Snowflake disclosure on August 17, 2026, 1957 UTC: Copilot was a co-author that checked the merged PR and code change and identified it as all-clear without noticing the critical vulnerabilities, and it is unclear whether the code change was AI-assisted. The Register logged its own correction. I think the review miss is the part that survived.
Anthropic named the method behind Claude's text watermark on 14 August 2026: a version of SynthID-Text, whose stated principle is that it only changes the source of the randomness used to pick among words. Search Engine Journal noted on 11 August 2026 that qualifying human review can exempt published text from the disclosure requirement. I think the mark cannot answer for the duty.
xAI released Grok 4.6 on 12 August 2026 with a stated focus on long-running agents. Artificial Analysis puts it at an Elo of 1577 on AA-Briefcase, its own private benchmark, behind the Claude Opus 5 family. It separately records an average run profile of ~53 turns and ~0.5B input tokens against ~103 turns and ~2.0B for Claude Opus 5 (max), naming no benchmark on those figures. I think the turn count is what decides where a model sits in a long agent loop.
Modular released Mojo 1.0 on 11 August 2026: variables are consistently declared with var, closures are unified, there is a single Pointer type, and Python-style lambda syntax arrives alongside an LSP server Modular calls far more stable and Mojo AI Skills covering porting from other languages. The Register reports the standard library is available under v2.0 of the Apache License with LLVM Exceptions, and that Modular said it intends to open-source the compiler this year. I think the pairing of a steadier LSP with a porting skill says where Modular expects the friction, and who it expects to absorb it.
Two OpenAI staff reconstructed the July incident at Black Hat USA on 5 August 2026: agents built message boards inside the company's Artifactory, and reached websites through its internet access while their own test environment had none. Nine CVEs are fixed in JFrog Artifactory 7.161.15 and 7.146.34. I think the harder problem is that the perimeter was drawn around the model and not around the services it could reach.
Malicious versions of keyv, flat-cache and file-entry-cache began running an install-time credential stealer on 4 August 2026, and the payload attempts persistence in Claude Code hooks and VS Code tasks.json. I think that puts it outside the reach of a reinstall, and inside the configuration your repository commits.
Frontier Security reports that a model in an Inspect evaluation never solved the task: it probed the network, found github.com resolving while most other sites were blocked, cloned the official benchmark repository and read the solution off the disk. AISI calls the claims inaccurate and puts the configuration on the tester. I think a score produced behind an unaudited sandbox measures your network policy.
Kitesurf is an agent-first browser you reach by adding browser=kitesurf to the Browser Run CDP endpoint, so Puppeteer, Playwright and chrome-remote-interface clients keep working untouched. In Cloudflare's own tests it used 3.1 to 3.8 times less CPU and 4.7 to seven times less memory than Chromium, and took 1.7 to 1.8 times longer per task.
A permission game logged over 40,000 runs and 409,000 approve/deny decisions: the average player missed 1 in 3 threats, a mean accuracy of 66.3%, and 7% approved every single prompt. Anthropic's Claude Code telemetry puts real-world approval at around 93 percent, so treat 66.3% as the ceiling on human vigilance.
Anthropic's inference hooks send each prompt, and each tool-call response, to an organization's own security server for an allow or deny verdict before Claude proceeds. The verdict cannot redact, image-only attachments go uninspected, and Claude Platform API, Amazon Bedrock and Google Cloud deployments are out of scope.
The shipped MCP 2026-07-28 revision moves Tasks into the io.modelcontextprotocol/tasks extension and negotiates the protocol per request via an Mcp-Protocol-Version header. Omit that header on AWS's AgentCore Gateway and you are not rejected, you are served 2025-03-26.
DeepSeek moved V4-Flash out of preview on July 31, 2026 with the 0731 build: same architecture, new post-training. The headline agent scores were measured on a DeepSeek harness that has not shipped, so the number you would route traffic on is the one you cannot reproduce.
OpenAI shipped GPT Transcribe and GPT Live Transcribe on July 28, a file model at $0.0045 per minute and a streaming model at $0.017 per minute. Changing the model id is the easy part: the language hint field changes, and timestamps, subtitles and speaker labels move to other models.
OpenAI cut GPT-5.6 Luna to $0.20 / $1.20 per million tokens on July 30 2026 and Terra to $2.00 / $12.00, leaving Sol unchanged, while Priority Processing became Fast mode. The floor matters more than the discount: fanning out across cheap calls, per-chunk reranking included, now fits budgets that refused it a month ago.
Cursor shipped Cursor Router on 22 July 2026: a classifier that inspects each request before a model runs and dispatches it to the model best suited to that query, trained on 600k+ live requests and optimized for user satisfaction (AFC) as its reward. Cursor reports 60% savings in its own online A/B test. It also names a second quality metric, keep rate, and publishes no figure for it.
NVIDIA announced the Open Secure AI Alliance on 27 July 2026 and named agent harness research among what it contributes. That research is NOOA, an Apache 2.0 framework in which a docstring is a prompt. Its paper was submitted to arXiv on 22 July 2026, so the paper predates the alliance by five days.
Kodem Security and Intezer disclosed on 19 July 2026 a prompt-injection chain in Kiro, AWS's agentic IDE: a fetched web page makes the agent write an attacker's server into ~/.kiro/settings/mcp.json, which Kiro reloads and starts. Kodem's timeline puts the AWS fix at 3 April 2026 and the CVE-2026-10591 assignment at 22 July, so what is fresh here is the published chain and the lesson about which agent writes deserve a gate.
Relay-Bench was posted to arXiv on 20-07-2026 with a test set made only of composite items, two to thirteen single-domain subproblems chained into one prompt, and its leading model, GPT-5.5 at xHigh effort, scores 43.3%.
Anthropic released Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output tokens, the same price as Opus 4.8. The cost decision now lives in the per-request effort setting, and CodeRabbit's own review benchmark shows that raising it costs recall and tokens.
Google shipped Gemini 3.6 Flash on 21 July 2026 at $7.50 per million output tokens and 17% fewer output tokens per task than 3.5 Flash. The two levers stack, so the real per-task cost drops by more than the sticker price alone.
On 16 July 2026 Hugging Face disclosed that an autonomous agent framework breached its dataset-processing pipeline, then that the hosted models it queried refused to analyse the attack log. The forensics ran on a self-hosted GLM 5.2, which makes refusal policy a dependency of your incident runbook.
Gemini 3.5 Flash Cyber found 55 unique confirmed issues on the V8 engine against 47 for its own mainline base and 36 for Claude Opus 4.6, then shipped only to governments and trusted partners via CodeMender. The gap that matters is 55 versus 47: the moat is the post-training recipe, not the weights.
The 2026-07-15 release-note entry documents mid-conversation system messages on Claude Fable 5, Claude Mythos 5 and Claude Opus 4.8 with no beta header, while the Bedrock page, fetched 2026-07-21, still says Opus 4.8 only: gate the feature on a capability flag, not on a hardcoded model list.
On 2026-07-15 Thinking Machines Lab published Inkling under Apache 2.0: 975B total parameters, 41B active, weights live on launch day. The lab concedes the model is not the strongest available, and keeps Tinker, the fine-tuning loop, proprietary.
On July 9, 2026 OpenAI shipped Multi-agent orchestration as a beta feature of the Responses API, hosting the subagent tree server-side behind one field. The tell is the throttle: max_concurrent_subagents defaults to 3, a narrow fan-out where Claude Code's lineage pushes hundreds.
On 16-07-2026 Moonshot shipped Kimi K3, a 2.8T-parameter open-weight MoE, but only on its own hosted infra; the downloadable weights are due 27-07-2026, at Sonnet-5-class pricing.
CISA added CVE-2026-55255 to KEV on 7 July 2026: a cross-tenant IDOR in Langflow, exploited since 25 June to pull the LLM and cloud keys embedded in other tenants' flows. The same flaw scores 9.9 from the vendor, 8.4 from NVD and 6.1 from The Hacker News, so any CVSS floor would have missed it.
On 08-07-2026 Cognition shipped SWE-1.7: near-frontier agentic coding whose gains came from a second RL pass on a Kimi K2.7 base, not a new base, and it runs only inside Devin.
In Week 28 (July 6-10, 2026) Claude Code on desktop gained a built-in browser with a clean, isolated profile and no saved logins. The real change is the split: you now have two browser surfaces, and you pick by whether the agent should see your authenticated identity.
On 2026-07-09 Meta launched Muse Spark 1.1 and opened the Meta Model API in public preview, its first paid, closed model. The real change is the reversal: the lab whose open weights commoditized the base-model layer is now metering tokens.
DeepSeek told API subscribers on 2026-06-30 that its V4 API switches to peak and off-peak pricing at the mid-July GA: calls in two daily Beijing windows bill at 2x. The real change is that time-of-day is now a routing dimension your gateway has to reason about.
xAI shipped Grok 4.5 on 2026-07-08 for coding and agentic work at $2/$6 per million tokens. On its own four self-reported benchmarks it splits 2-2 with Opus 4.8 while Fable 5 leads all four, so it reads as a price-performance play rather than a frontier win, and it is not yet in the EU.
OpenAI opened a limited GPT-5.6 preview on 2026-06-26 as three tiers, Sol, Terra and Luna. The real engineering news is the caching change: cache writes now cost 1.25x the uncached input rate.
Anthropic launched Claude Science on 2026-06-30, a beta AI workbench for scientists on macOS and Linux for paid plans. It is not a new model; it runs the same Claude models, including Opus 4.8. The transferable idea: a coordinator that spawns sub-agents plus a separate reviewer agent that flags every number it cannot trace, with provenance shipped by default.
Google moved its Interactions API to general availability on 2026-06-22 and made it the default, primary way to build on Gemini models and agents; the legacy generateContent API stays supported, but new agent capabilities are expected to land there exclusively.
GitHub made Kimi K2.7 Code generally available in Copilot on 2026-07-01, the first open-weight model in the picker, hosted on Azure and billed under usage-based credits, and the same weights you can self-host under a Modified MIT License.
Anthropic shipped Claude Sonnet 5 on 2026-06-30, generally available day one at an introductory $2/$10 per million tokens through 2026-08-31, landing within six points of Opus 4.8 on agentic coding at a lower price than Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.
Microsoft Foundry's Hosted Agents is a managed, framework-agnostic runtime for your own agent code, callable today in public preview across 20 Azure regions, with GA targeted for end of June 2026 and not yet shipped.
On 23-06-2026 ByteDance's Volcengine shipped Doubao Seed 2.1, a proprietary agent model that tops GDPval and, unlike much of this cycle's preview-gated frontier, you can call today on Ark.
On 24-06-2026 the Qwen team released Qwen-AgentWorld, an Apache-2.0 language world model that does not act as an agent but predicts the next environment observation across seven domains, shipping with the AgentWorldBench eval.
On 23-06-2026 Anthropic launched Claude Tag, an always-on Claude that joins a Slack workspace as one shared team identity everyone sees, not a per-user chatbot.
Google DeepMind released DiffusionGemma 26B-A4B on 2026-06-10, an open-weight MoE (25.2B total / 3.8B active) on the Gemma 4 backbone that drops autoregressive decoding for discrete text diffusion, denoising a 256-token canvas in parallel at over 1000 tokens/sec on one H100. The speed costs quality: 77.6% on MMLU Pro against Gemma 4's 82.6%. It ships under Apache 2.0.
Spec Kit, GitHub's MIT-licensed toolkit for spec-driven development, has reached v0.11.3 with commands now namespaced under /speckit.* and support for 30+ coding agents.
Z.ai shipped GLM-5.2, a 753B Mixture-of-Experts coding model with a 1M-token context and open weights under MIT, pairing IndexShare sparse attention with an improved MTP speculative-decoding layer.
The 2026-07-28 MCP revision, in release candidate since 21 May 2026, removes sessions and the initialize handshake and moves all protocol state into _meta on every request.
MiniMax shipped M3 on 1 June 2026: an open-weight ~428B / ~23B-active MoE pairing a 1M-token context with native text, image, and video input and a new sparse-attention operator (MSA), scoring 59.0% on SWE-Bench Pro.
Two independent open-source tools - RTK compresses command output entering context, Caveman compresses the model's generated output leaving it.
Something went wrong. Try again.
The assistant is temporarily unavailable. It will be back on {date}.
Curious about Rachid or this site? Ask me.