AI Dev Weekly #31: Claude Duet, Beam, Mercury 3 and Longer Prompt Caching

AI Dev Weekly is a Thursday series covering the AI developer news that changes what you can build, what it costs, and how you should operate it.

This week brought a useful split in the AI market. Anthropic wants an agent to keep working between your conversations. Reflection and Meta released substantial models you can actually download. Inception is competing on the time you spend waiting. OpenAI and GitHub, meanwhile, shipped changes to the economics and safety of the machinery around those models.

The headline is not another universal winner. It is a wider set of deployment choices, each with different access, hardware, pricing, and permission boundaries. A research preview is not general availability, open weights do not imply an Apache license, and a fast output stream is not a completed task.

This edition covers developments through October 22, including October 15 releases that landed after the previous issue. It is based on official announcements and documentation, not hands-on benchmarks of the newly released products.

1. Claude Duet joins the proactive-agent race

Anthropic introduced Claude Duet on October 21 as a research preview with a waitlist, starting with US Pro and Max subscribers. It is a new proactive-agent app, not a new model API. The product uses Claude Opus 5.5 and can continue work on a virtual machine with its own browser while your devices are off.

Duet is available through standalone iOS and Android apps and claude.ai. It carries over account memory, skills, and connected apps. During the preview, it shares your existing plan's usage limits. Anthropic says connected accounts are checked every few hours, rather than continuously. See the Duet announcement.

That last detail is important. A proactive assistant is not necessarily a real-time monitor, and background activity is not free simply because you are not watching it. Teams evaluating this kind of product should separate three questions: what can it observe, what can it decide, and what can it change?

My take: Start with work that produces something reviewable: a meeting brief, a research note, or a proposed plan. Be much more deliberate about anything that sends messages, changes appointments, or acts through an authenticated browser. The useful test is whether the agent asks for judgment at the right moment, including when nothing should happen.

This continues the shift we covered in OpenAI Dots, but access and product controls are not interchangeable. For our Belgian and wider European readership, a US-first waitlist is not a local rollout. Our agent memory patterns and production agent checklist help frame the state and authority questions behind the product demo.

2. Beam and Muse Spark Open turn announcements into downloadable weights

Reflection released Beam's weights on October 16, following its earlier announcement. The release includes BF16, FP8, and NVFP4 checkpoints, a speculative-decoding drafter, and an Apache 2.0 license. Reflection also released Mirror CLI, with tool approvals, resumable conversations, and support for isolated worktrees. The availability announcement describes the model and surrounding deployment tooling.

Meta's Muse Spark Open arrived on October 15, after last week's issue. It is a different release from the hosted Muse Wonder family covered in edition #30. Meta distributes its weights under the Muse Model License with limited commercial terms, not Apache 2.0. The Meta announcement also describes running it across two M5 Ultra Mac Studios, which should not be confused with a claim that it fits a typical laptop.

The model references make the differences concrete:

Both have million-token-class context windows. Beam's model card identifies it as text-only and reports results separately for its numerical formats. Meta's deployment documentation explains that the active parameter count controls only part of the compute picture: serving still requires loading the full weights. KV cache and runtime overhead add to the footprint.

My take: These are meaningful releases for teams that want deployment control, but neither is a small local coding model. Compare the actual checkpoint, serving engine, memory requirement, concurrency, and license before comparing benchmark scores. A hosted evaluation can establish whether the model is interesting without first buying the hardware to serve it.

The next useful step is a fixed repository task with known tests and a documented tool set. Keep quantization and harness settings in the result, because changing either can change the outcome. Our GPU memory planning guide covers capacity, while the open-weight coding model guide helps structure a broader shortlist.

3. Mercury 3 puts diffusion back into the latency conversation

Inception launched Mercury 3 on October 19. Its announcement describes a diffusion-based language model with a 1M-token context window and reports 1,068 tokens per second on NVIDIA GPUs. That is a vendor-reported speed, not a measurement from this site or a promise about your application.

Regular pricing is $0.20 input and $0.70 output per million tokens. The launch offer halves those rates to $0.10 and $0.35. Keep the promotional and regular prices separate in your budget. Inception lists access through its own platform and partners including OpenRouter, Baseten, and Vercel, with OpenAI-compatible API integration.

The relevant question is not whether a demo can print text quickly. It is whether faster model work makes the user-visible job finish sooner. An agent can spend most of its time waiting on a browser, fetching documents, running tests, or recovering from a bad tool call. None of those delays disappears when output generation accelerates.

My take: Test Mercury on a latency-sensitive workflow where the model is a known bottleneck. Record time to first useful result, total completion time, tool failures, retries, and task acceptance. Do not use output throughput as a substitute for all five. Compatibility also deserves testing: accepting a familiar request shape does not establish that every streaming event or tool behavior matches another provider.

We covered the preceding generation in the Mercury 2.5 explainer. Treat that as architecture background, not a current price sheet. Our API pricing comparison and fallback patterns cover the surrounding buying and reliability decisions.

4. OpenAI adds 12-hour caching, with a higher write rate

OpenAI added 12-hour prompt caching on October 20, according to the API changelog. For GPT-5.6 and later in the Responses API, the relevant control is prompt_cache_options.ttl.

The caching documentation distinguishes the two options:

These are total write rates, not an extra charge added on top of the normal input rate. Read rates stay the same across lifetimes. The longer option is a minimum cache lifetime, not a deletion deadline. Older-model retention controls are separate.

My take: Longer retention is useful for stable instructions, tool definitions, or reference material reused after a long gap. It is less compelling for a prompt prefix that changes on almost every request. A higher write rate only earns its keep if later cache hits recover the cost.

Before enabling it broadly, replay a representative request trace and count real prefix reuse. Separate stable shared material from rapidly changing session state. Then inspect billed writes and reads instead of assuming repeated-looking prompts always hit the cache. Also keep caching separate from application memory: cheaper repeated context does not create durable task state or correct recovery after an interruption. Our prompt caching explainer and retrieval versus memory versus tools guide explain those boundaries.

5. GitHub's package firewall targets installation time

GitHub introduced its package firewall in public preview on October 20, with an October 21 clarification to the announcement. It currently supports npm packages only, is available to GitHub Enterprise Cloud customers including data-resident enterprises, and is free during preview. GitHub says GA will use metered billing.

The firewall sits between configured endpoints or internal registries and upstream repositories. It checks metadata and downloads against enterprise policies, including cooldown periods, advisory information, and supply-chain signals. Agents and CI hosts need to route their package requests through it. See the package firewall announcement.

This addresses a different layer from last week's workflow lockfiles. Locking establishes which revision should run. An install-time gate decides whether a requested package should be allowed under current policy. Neither establishes that all code in an agent's workspace is safe.

My take: This is worth a controlled trial if coding agents routinely install dependencies in your enterprise. Verify the denied path, exception review, logs, and behavior when the service cannot be reached. Do not assume a package-manager setting alone prevents a determined process from reaching an upstream registry by another route. Network controls and scoped credentials still matter.

Start on disposable workspaces with a small policy. Confirm that ordinary builds work and that a blocked dependency produces a clear, reviewable failure. Our agent sandboxing guide and AI supply-chain risk guide cover the wider containment model.

Quick hits

• OpenAI models are entering Google Cloud, but not as a broad public launch. Google's October 21 announcement describes private preview for selected enterprise customers, including GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna. Marketplace purchasing and cloud commitments are part of the offer. Antigravity access, Confidential Compute, and ZDR options are described as coming later, not already available everywhere.
• MiniMax M3.1 Flash launched October 19. The release post lists $0.15 input, $0.60 output, and $0.015 cached input per million tokens. It reports 73.7% on DeepSWE v1.1, a vendor result rather than our evaluation. API, M Plan, and MiniMax Code access are available; weights are promised later.
• Claude Security moved further into ongoing maintenance. The October 21 update adds continuous repository scanning, tracking of landed fixes, project threat-model context, and API/MCP connections. It remains a public beta for Enterprise customers. Finding issues and proving that fixes have shipped are separate jobs; the latter is the useful operational shift here.
• GPT-Live gained prompt-generated voices. OpenAI's October 16 API update makes prompt-to-voice available to all API customers. Creating the voice is free; subsequent speech generation uses normal session billing. The same release week also added document-creation tooling to hosted Agents API sandboxes and service-tier controls for projects.

What I would do this week

Pick the bottleneck before picking the release. If you need deployment control, evaluate one open-weight checkpoint. If waiting on generation dominates your workflow, test Mercury. If repeated context dominates spending, measure cache reuse. If agents install packages, inspect the installation boundary before giving them more autonomy. The best upgrade is the one that improves an observed problem without quietly expanding authority.

FAQ

Can European users access Claude Duet now?

The announced research preview starts with US Pro and Max subscribers through a waitlist. Do not assume a European launch or immediate access for every paid account.

Are Beam and Muse Spark Open under the same license?

No. Beam uses Apache 2.0; Muse Spark Open uses Meta's Muse Model License and associated terms. Download availability does not make those licenses interchangeable.

Does 42B active mean Muse Spark Open fits in 42B-model memory?

No. Active parameters describe routing during inference. The full checkpoint and serving overhead still determine the memory footprint.

Is the 12-hour cache cheaper for every application?

No. It has a higher write rate. Savings depend on actual later reuse, and its minimum lifetime should not be read as a data-deletion guarantee.

Does GitHub's firewall protect every package ecosystem?

Not in this preview. The announced support is npm, and protection depends on routing the relevant requests through the configured service.

Originally published at https://www.aimadetools.com