|
Happy Saturday, there 👋
Anthropic shipped a new flagship yesterday. I skipped the benchmark table and opened the migration guide first, because that is where a launch quietly tells on itself. Two screens in, there it was: thinking is on by default now, and turning it off at the top effort levels makes the API answer with a 400.
My bias, up front. A launch-day scorecard is a lab result on a good day, and I read it as exactly that. What changes my Monday is the migration guide, so this issue starts there and works outward through a week when the ground under your agents kept shifting.
The Big Thing
Anthropic shipped Opus 5, and the upgrade is not a drop-in
Anthropic launched Claude Opus 5 on Friday, and the pitch is blunt: cheaper than Fable 5, less restricted than Fable 5, and ahead of Fable 5 on a chunk of the benchmarks in the announcement. It runs about $5 in and $25 out per million tokens, roughly half of Fable 5's input price. For a team renting a frontier model to run agents, that math by itself is a reason to look.
Then you read how to call it. Thinking is adaptive by default now. The old switch, thinking set to enabled, comes back a 400 with an invalid_request_error, and you move to adaptive thinking plus an effort level of low, medium, or high. Turn thinking off at the top effort levels, xhigh or max, and you also get a 400, a real break from Opus 4.8, where those two settings were independent. None of this is in the launch copy; it surfaces the first time you swap the model name and ship without touching the thinking parameters.
The capability jump is real, and it shows up where agents live. On Anthropic's announcement, Opus 5 climbs on agentic coding and moves from the mid-50s to about 70 on OSWorld computer-use. The number I keep staring at is Zapier's AutomationBench, where Anthropic reports Opus 5 moving from 17 to 26. That is a real jump, and it still leaves a model that fumbles roughly three of every four end-to-end automations. Third-party leaderboards put it near 96% on SWE-bench Verified. Read the launch numbers as the vendor's homework, and take the leaderboards as early signal.
Anthropic was also straight about the edges. Opus 5 stays fenced on offensive security: it will hunt for bugs in source code, which reads as defense, but it will not scan a compiled binary for you. The company says its safety classifiers should fire about 85% less than they do on Fable, and it shipped an opt-in beta, Automatic Fallbacks, that reroutes a blocked request to a weaker model so you get an answer instead of a wall. On the hardest cyber work it still trails Anthropic's Mythos 5.
Ship it? Move your agent and coding workloads over this week, once the call is fixed. Set the effort level on purpose, swap enabled thinking for adaptive, and do not disable thinking at the top effort levels. Re-run your evals before production, because cheaper and smarter is still a behavior change under the hood. If refusals bite mid-run, flip on Automatic Fallbacks.
Anthropic Opus 5 announcement · TechCrunch · CNBC · Claude migration guide
Tour de Headlines
🐝 Block gave every AI agent a passport. On July 21, Jack Dorsey's Block shipped Buzz, a free, open-source workspace that is part team chat and part code host, built for people and agents to work side by side. The trick sits underneath. It runs on Nostr, so every agent joins with a cryptographic keypair of its own, and every message, review, workflow step, and Git event becomes a signed record in a hash chain. When an agent acts, the log can prove which one and what it touched. Buzz runs agents built on Claude Code, Codex, and Block's goose, the on-machine agent we handed you as a Tool back in June. Desktop now, mobile soon. Ship it? Try it if agent accountability is the part you keep bolting on late.
TechCrunch on Buzz
🏭 Hitachi says agents made it 240x faster, then aims for 30%. Also on Friday, Hitachi put out its Agentic AI Integration Platform, which drops task-specific agents into every stage of building a system, from requirements through design, code, testing, and operations, the large, high-stakes kind it sells to banks and railways. Internal runs claim up to 240 times the productivity in the requirements phase and around 200 times through design and test. Read those as best-case lab numbers under chosen conditions, because the goal Hitachi will stand behind in public is a 30% lift by fiscal 2027. The direction is the signal, the multiple is marketing. It leans on Anthropic, Google Cloud, and OpenAI underneath, so the bet here is a systems integrator wiring agents across the whole build. Ship it? Watch.
IBTimes on Hitachi
⏰ Following up: MCP goes stateless Tuesday. Back in June we led with this clock, and it lands on Tuesday, July 28. The stable spec pulls session state out of the protocol core: the initialize handshake goes, the Mcp-Session-Id header goes, and servers stop holding a session per client. That makes horizontal scaling easier, since a request can hit any box with no sticky affinity. The spec adds a formal deprecation policy for the pieces it is retiring, but the new SDKs assume stateless right away, so the audit is a this-weekend job. Ship it? Audit your session-dependent servers before Tuesday.
MCP release notes
|
Sponsor
Measure the human signal in every conversation.
Your agents get scored on every call they handle. RapportScore does the same for your people, reading the communication signals in real sales and support conversations so teams can see what lands and adjust before the next one. Built by the crew behind this newsletter.
See your team’s score →
|
Tool of the Day
⚡ Anthropic Automatic Fallbacks
One opt-in setting so a safety refusal returns a routed answer instead of a 400.
Tucked inside the Opus 5 launch is a setting worth flipping by itself. Turn on Automatic Fallbacks, in beta, and any request that trips Anthropic's safety classifier gets rerouted to a less powerful model that will answer, rather than handing your agent an error. It exists because the guardrails have started getting in the way of legitimate work, offensive-security research most of all, and a stalled agent mid-run is a failure too. This is a graceful degrade, not a jailbreak: you trade some capability for a response you can keep working with. Ship it? Turn it on if your agents ever stall on a refusal.
Read the Opus 5 announcement →
Worth a Click
- Before you touch a server Tuesday, read this. The Register’s plain walk-through of what the stateless switch removes and what it changes about the way you run MCP. The Register on MCP going stateless
- Why the safety net is snagging real work. The reporting behind today’s Tool: guardrails are slowing legitimate offensive-security research, and vendors are now building escape hatches around the classifiers they shipped. TechCrunch on guardrails and security work
This week's theme was bigger than any single launch: the ground shifted under the way you call your agents. Opus 5 changed the default, so thinking is on and effort is a dial you now have to set. MCP is about to pull session state out of the protocol. Buzz is handing every agent a signed identity. Three different layers, model, protocol, workspace, all rewiring their defaults inside the same seven days, and none of them waited for your go-ahead.
So the discipline is small, and it keeps coming back. When a default changes, read the docs first, before the 400 shows up in your logs. Give your agents identities and audit trails while the tooling to do it is finally arriving. Treat every drop-in upgrade as a behavior change until your evals say otherwise. The models keep getting cheaper and louder. Your job is to keep knowing exactly what you asked for.
Ron
|
