AI Builders Digest

Opus 5 ships at Opus 4.8 pricing, with a 2.5x Fast mode and top alignment scores

Opus 5 is out on every paid plan and the Claude API at the same price as Opus 4.8, default on Max and the strongest model on Pro. Fast mode runs it at roughly 2.5x the default speed. Anthropic's automated behavioral audit calls it the most aligned model they have shipped, with the lowest rates of reckless or deceptive behavior. On cybersecurity it beats Opus 4.8 but stays substantially behind Mythos 5 at developing exploits, with safeguards tuned to permit vulnerability fixing while blocking high-risk use.

• #products
• #evals

Every's Opus 5 verdict: delete your skills or the model fights you

After a week of testing across coding, writing, and their internal agent, Every's read is that Opus 5 is a hard model to love. It argued with instructions, stopped before finishing work, and broke their existing skills and plugins badly enough that the first reaction was "What have they done to my boy?" Deleting those skills and rebuilding from scratch flipped the result, and lower thinking effort worked better than high, since more reasoning time surfaced more of the annoying behaviors. Shipper's placement: the personality of a genius model without the top end, so it sits awkwardly between Fable for hard problems and GPT-5.6 for everything else.

• #products
• #evals

Box benchmarks Opus 5: +30% life sciences, +17% due diligence

Box ran Opus 5 through its Complex Work Eval, an agentic benchmark of real enterprise document work, and posted gains across every vertical: life sciences +30%, technology +19%, due diligence +17%, healthcare +13%, legal +12%. The pattern behind the numbers matters more than the numbers: Opus 5 stayed thorough as checklists grew instead of catching the obvious items and stopping, and correctly handled a contract exception that Opus 4.8 mis-scored. Levie also signed Box onto the open weights letter, arguing open and closed is not zero sum, since post-training for finance, legal, or life sciences gives hundreds of shots at a vertical instead of waiting on a few frontier labs.

• #evals
• #agents
• #open-source

Anthropic says layered defenses push prompt injection success to near zero

Buried in the Opus 5 system card, and more interesting than any eval score: it is Anthropic's least prompt-injectable model yet, holding up across injection evals and red teaming. Stack model alignment with injection probes and Claude Code's Auto Mode and the attack success rate drops to roughly zero.

• #security
• #agents

Anthropic: contain agents at the environment layer, because approvals fail

Anthropic's engineering team documented how it sandboxes Claude across claude.ai, Claude Code, and Cowork, and the honest parts are the failures. Telemetry showed users approved about 93% of permission prompts, so per-action consent quietly stopped being oversight; an OS-level sandbox cut prompts 84% instead. In an internal red team, an employee was phished into pasting a prompt that asked Claude to read ~/.aws/credentials and POST them out, and it succeeded 24 times in 25 tries, because when the user types the instruction there is nothing anomalous for a classifier to catch. A third-party disclosure showed the reverse failure: the egress allowlist passed traffic to api.anthropic.com, so an attacker's embedded API key uploaded workspace files to their own account. The recurring lesson is that hypervisors, seccomp, and gVisor held, while the custom proxy Anthropic wrote itself is what broke.

• #security
• #agents

Claude Code cut about 80% of its system prompt for the newest models

Anthropic removed roughly 80% of the Claude Code system prompt for the Claude 5 family and wrote up what that taught them about system prompts, skills, and CLAUDE.md files. Thariq's pairing suggestion: Opus 5 as the daily driver, Fable for planning, brainstorming, and the hardest bugs.

• #agents
• #products

Opus 5 makes consultant-grade decks and spreadsheets, and does it token-cheap

Six months on, Albert says Opus 5 produces near-superhuman spreadsheets and slide decks at the level a consultant would deliver. A lot of the launch work went into token efficiency across domains while still raising the intelligence bar, and he prefers it over Fable 5 for many coding tasks.

• #products
• #evals

DoorDash built its own delivery robot because nobody was building for its use case

DoorDash has been working on autonomy since 2018, starting as one and a half engineers of skunkworks, and spent years partnering with sidewalk robot and robotaxi companies before concluding it had to build the vehicle itself. The reason is a use-case gap: a 2 mph sidewalk robot cannot serve a three to five mile delivery, and a 4,000 pound robotaxi with chairs and AC is absurd for carrying a couple of burritos, so Dot is a 300 pound, 20 mph bike-lane vehicle that has been running L4 deliveries in Phoenix for two years. Stanley Tang's argument is that the hard part was never the model: booting hundreds of robots each morning with a hacked-together Jenkins script, regen braking that overpowers the battery, torque differences when two wheels sit on leaves, and finding which front door a GPS pin actually means. DoorDash claims the unique asset is 10 billion deliveries of drop-off data that does not exist in Google Maps. Andy Fang added that Ask DoorDash drives 50% of restaurant sessions toward places the user has never ordered from and 40% larger grocery baskets, and that internal AI spend rose 20x from January to June before flattening. Tang's counterintuitive prediction: in ten years DoorDash will have more human Dashers, not fewer, because demand grows faster than any single modality can absorb.

• #agents
• #hardware
• #products

Masad presses Anthropic to state its position on open weights

Masad is publicly asking whether Anthropic will sign the open weights letter, and suggesting Anthropic employees ask leadership to clarify whether the company favors banning open weight models. Separately he needled VCs who passed on Etched in early rounds, noting the early believers now take less dilution.

• #open-source
• #policy
• #funding

Altman backs open source and teases a pro-ultra-superhard setting

Altman wants the US to win in AI on both open source and proprietary models, and says he is glad to see the open weights push. He also asked for feedback on something called pro-ultra-superhard.

• #open-source
• #policy

Tan: AI productivity gains will take 10 years, not 2

Macro productivity gains require managers and CEOs to greenlight radically different staffing and workflow plans, and Tan does not think they have done it yet. His framing for why this matters: how fast a country adopted new technology over the last 200 years explains at least 25% of why some nations are rich today, and the Ottomans got the printing press eventually, but eventually was expensive.

• #policy
• #adoption

The scarce AI skill is adapting foundation models to messy real workflows

Guru sees a large multi-year opening for people who can take real-world workflows and adapt foundation models to them, which means understanding how work actually gets done, designing evals, doing post-training, and building feedback loops that keep improving the model. That is how a general-purpose model becomes exceptional in a specific domain, and today the skillset sits inside a handful of labs.

• #evals
• #talent

Speed, not intelligence, is the bottleneck now

Zhang's ask of any model right now is speed, because intelligence is already good enough. A one to five minute wait per task is the worst possible window, too short for deep work and too long to just watch, so the wait gets filled with scrolling. Her related observation: once agents sit in your chat groups and meetings, transcripts become PRDs, which levels the field for verbal communicators in a work culture that has always rewarded writers.

• #agents
• #products

Model routing consolidates, with Stripe rumored to buy OpenRouter for $10B

Turck counts a big week for routing: a rumored $10B Stripe acquisition of OpenRouter, Cursor Router on Wednesday, Runway Router the next day, on top of existing routers at Databricks, Vercel, Cloudflare, Dataiku, AWS, and Google, all meaning fairly different things under one word. His other note is the irony of top AI researchers building recursive auto-research and researching their way out of their own jobs.

• #funding
• #products

ChatGPT Work goes global across every paid plan

ChatGPT Work is now available worldwide on all paid plans across mobile, web, and desktop, and is likely already installed on your phone.

• #products

Gemini Spark goes live for US Google AI Pro subscribers

Gemini Spark is live for all Google AI Pro subscribers in the US, with global expansion next. Woodward's framing is action over conversation: drop in a school calendar PDF, tell Gemini to add every No School day to Google Calendar, done.

• #products
• #agents

Pure software is getting hard for indie developers to monetize

Yang agrees that software alone no longer supports indie developers, and that you now need software plus something else such as services. He is also running Codex by talking to ChatGPT Voice from bed, which works well but requires remembering the names of all your long-running threads.

• #products
• #agents

An autoreview skill ran 66 rounds on a single refactor

Steinberger hit a new record with his autoreview skill: 66 review rounds on one gnarly refactor.

• #agents

SmolForge adds customizable skins and spritesheet animations

SmolForge picked up four new features, including customizable skins and spritesheet animations. Swyx is also building a new gsuite, motivated by defaults in the existing one that he considers indefensibly bad.

• #products