Sonnet 5 Hits Make

Make added Claude Sonnet 5 to its platform, exposing Anthropic’s newest Sonnet-tier model across three distinct integration surfaces. The headline is a much larger context window. The footnote is a tokenizer change that quietly raises your costs.

What shipped

Sonnet 5 is now available through three surfaces inside Make:

• Make AI Toolkit, usable without an API key if you route through Make’s own AI Provider.
• The standard Anthropic Claude module, which requires your own Anthropic API key.
• Make AI Agent (New), Make’s agentic automation runtime, also via your Anthropic key.

The headline capability is a 1,000,000-token context window, paired with what Make describes as improved reasoning and better cost control for high-volume scenarios. Output is configurable up to 128,000 tokens, which Make frames as a guardrail for recursive or scheduled runs that could otherwise blow past intended limits.

There is one significant footnote operators need to internalize before flipping the switch. Sonnet 5 ships with a new tokenizer, and Make explicitly warns that the same input text will produce roughly 30% more tokens than it did under Sonnet 4.6. Existing scenarios will consume more credits per run and may hit output truncation earlier than they did on the prior model.

1Minput context tokens

128Kmax output cap

+30%tokens per input vs 4.6

Key details

• Model: Claude Sonnet 5, from Anthropic.
• Context window: 1,000,000 tokens on input.
• Output cap: configurable, up to a maximum of 128,000 tokens.
• Tokenizer change: same input text produces about 30% more tokens than under Sonnet 4.6.

Make’s stated benefits are large single-call input handling for documents, PDFs, and exports, multi-step planning and tool use inside agent scenarios, predictable recurring cost at Sonnet tier, and configurable output caps as guardrails for recursive or scheduled scenarios.

Why it matters

Make sits in an awkward middle layer of the AI tools stack. It is not a model provider, and it is not an IDE. It is the connective tissue non-engineers use to wire models into real business workflows. That makes Make’s model-availability announcements less about the model itself, since Sonnet 5 already exists wherever Anthropic ships it, and more about how quickly the automation layer catches up to the frontier, and on what terms.

The interesting move here is the three-surface rollout. Make is not just adding Sonnet 5 to the Anthropic module. It is exposing the same model through its own managed AI Provider, where no API key is required and Make handles the plumbing, and through Make AI Agent, its agentic runtime. That mirrors the direction most automation platforms are pushing: a low-friction default for simple calls, BYO-key for teams with existing Anthropic contracts, and a separate surface for multi-step agent work where tool use and planning behave differently than single-shot completions.

The context window is the operator-relevant piece

Automation scenarios have historically been the worst possible place to do chunking, because splitting a document across modules means splitting state across a scenario graph, which is expensive to debug and brittle to maintain. A million-token window inside a single Anthropic module collapses a whole category of scenario complexity into one call. For document-heavy workflows like contract review, PDF exports, and long transcripts, that is the right pitch.

The tokenizer footnote deserves more attention than Make’s docs give it. A 30% token inflation on identical input text is a silent capacity change for anyone with scenarios pinned near output caps, credit budgets, or context ceilings.

Sonnet 5's new tokenizer counts roughly 30% more tokens for the same input, so scenarios sized against Sonnet 4.6 will cost more and may truncate earlier. Re-measure before leaving anything on autopilot.

How it compares

Sonnet-tier availability inside automation platforms has become table stakes. The differentiation is on the agent-runtime side. Make is pushing Sonnet 5 into its Make AI Agent surface at the same time as the standard module, which signals the platform is treating agentic scenarios as a first-class execution mode rather than a bolt-on. Most no-code automation tools still expose Anthropic models as single-shot completion nodes and rely on the builder to hand-wire branching logic across modules. A dedicated agent runtime with planning and tool use is a different shape of primitive.

The 1M context ceiling is the same ceiling operators will find in Anthropic’s direct API for long-context Sonnet configurations, so Make is not offering a smaller window than the source. What Make offers that a raw API call does not is the ability to feed that window from other Make modules, like a Google Drive fetch, a CRM export, or an email thread, without an intermediate assembly step.

Where Make is quiet: no benchmark numbers, no comparison to Sonnet 4.6 on reasoning quality beyond the word “better,” and no guidance on which of the three surfaces is preferred for which workload.

Who should care

• Document-processing operators running contract intake, PDF-to-structured-data pipelines, or long-transcript summarization should re-plumb the chunking logic out of their scenarios. If you built a splitter-and-reducer pattern to work around Sonnet 4.6’s context limits, that pattern is now dead weight in a lot of cases. Test a single-module version.
• Teams running scheduled or recursive scenarios on Anthropic modules need to audit credit consumption this week, not next quarter. The 30% tokenizer inflation is silent and cumulative. Anything on a nightly cron with a monthly credit budget is the highest-risk surface.
• Builders on Make AI Agent (New) should treat Sonnet 5 as the default candidate for multi-step decision-making workflows, given Make’s own framing around planning and tool use. If you were routing agent decisions through a smaller model to control cost, re-evaluate whether Sonnet 5 collapses that into one call.
• Teams without an Anthropic contract who want to try Sonnet 5 without procurement overhead can use the Make AI Toolkit path with Make’s AI Provider as the no-key entry point.

What to watch next

The most immediate thing to watch is credit-consumption complaints. A 30% tokenizer inflation announced in a docs footnote is the kind of change that produces support tickets three weeks later when someone’s monthly usage report lands. Expect Make to either surface a tokenizer-aware credit estimator in the module UI or publish a clearer migration note within the next few weeks.

Second, watch whether Make AI Agent (New) becomes a serious agentic runtime or stays a side surface. Shipping Sonnet 5 into the agent module on day one signals that Make wants agentic automation to be a real product line rather than a marketing checkbox. The test is whether the agent module gets its own feature velocity, like memory, tool registries, and evaluation, or whether it stays locked to whatever the standard Anthropic module does.

Third, watch the 1M-context claim under load. Long-context Sonnet performance in production has historically degraded on retrieval-style tasks well before the theoretical ceiling. Operators moving from chunked scenarios to single-payload scenarios should A/B the two patterns on their actual data before decommissioning the chunker.

Finally, watch for Make to add other frontier models to the same three-surface pattern. If Toolkit, standard module, and Agent (New) become the default rollout template, that is the shape of how Make plans to compete with more code-first orchestration layers.

Open questions

Make’s announcement leaves several operational questions on the table.

1. What does “better reasoning” mean in numbers? No benchmarks, no comparison table, no task-level examples of where Sonnet 5 beats Sonnet 4.6. Operators picking between the two are working from vibes.
2. Is Sonnet 4.6 being deprecated, and on what timeline? The announcement adds Sonnet 5 but says nothing about the fate of prior Sonnet versions, prompt-caching compatibility, or scenario migration paths.
3. How does the no-key AI Provider path meter usage? Does it apply the same 30% tokenizer inflation to credit consumption that BYO-key users will see on their Anthropic bill? The announcement flags the inflation but does not specify which billing surface absorbs it.
4. Does Make AI Agent (New) expose the full 1M context, or is agent-mode context reserved for planning state and tool-call transcripts with a smaller effective input budget? Agentic runtimes often carve context differently than single-shot modules.
5. Are there rate limits or concurrency caps specific to Sonnet 5 inside Make that differ from other Anthropic models? Not addressed.

Source: help.make.com