Make Review 2026: AI Agent Fallbacks, Prompt Caching, and MCP Audit Logs

Make's October 9 release is aimed at what happens after an AI agent is built: it keeps running when the model provider fails, costs less on repeat calls, and leaves an audit trail an admin can query from an MCP client (help.make.com, Oct 9 2026, retrieved 2026-10-13). The rating holds at 4.5. For anyone running agents in production on Make, turn on a fallback connection this week; it is free on every plan and it is the change most likely to save a run.

September's headline was the Make plugin for ChatGPT, a new front door. October is about operating what you already built. Of the October 9 changes, fallback connections and prompt caching matter to almost every AI Agent user, SSO routing and audit-log MCP tools matter mostly to IT and security, and the rest is a model refresh. One finding from reading the docs side by side changes who should care about the governance half: Make's own audit-log page says audit logs are "available on the Enterprise plan only" (help.make.com/audit-logs, retrieved 2026-10-13), even though the new MCP tools that query them are listed for all plans.

What changed since September 10

Each row compares the September 10 version of this review with what Make shipped through October 12.

Axis	September 2026 baseline	October 2026	Source
Agent provider failure	Run fails	Retries once on a fallback connection; Make AI Provider is the default fallback (Oct 9)	help.make.com, Oct 9
Repeat agent calls	Full prompt reprocessed each call	Prompt caching for instructions and tool definitions (Oct 9)	help.make.com, Oct 9
Audit trail access	Dashboard filters only	Read-only MCP tools to list, filter, and inspect audit events (Oct 9)	help.make.com, Oct 9
Social sign-in on a company domain	Could bypass the org's SSO flow	Routed into the org's SSO flow with email pre-filled (Oct 9)	help.make.com, Oct 9
Newest models	Claude Fable 5.1, Gemini 3.7 Flash	Plus Claude Sonnet 5.5, GPT-6.1 Sol, Claude Haiku 5.5 (Oct 9, Oct 12)	help.make.com/2026
Platform event alerts	Build your own monitoring	Signed event subscriptions with a delivery log, Teams and Enterprise (Oct 8)	help.make.com, Oct 8
Pricing	Core 9 USD / 10,000 credits	Held	make.com/en/pricing

Two smaller items round it out. On October 7 Make redesigned the scenario toolbar so Run once and Save sit side by side, and added a Save as version button that stores a description with each version, so version history reads like a changelog (help.make.com, Oct 7, retrieved 2026-10-13). On October 12 an app-update roundup added Claude Haiku 5.5 and set a Wealthbox OAuth deadline of October 21 for anyone still on API-token connections (help.make.com, Oct 12, retrieved 2026-10-13).

Fallback connections: what they catch, and where they stop

You set a fallback connection in the Make AI Agent module under advanced settings. If the primary connection fails, Make retries the run once on the fallback (help.make.com, Oct 9, retrieved 2026-10-13). Make's feature note lists what it is meant to absorb: temporary provider outages, timeouts, rate limits, and invalid or expired connections (community.make.com, retrieved 2026-10-13).

Expired connections are the case most teams will hit first. An expired API key on your own Anthropic or OpenAI account used to fail every agent run until someone noticed. With a fallback set, the run moves to Make's AI Provider (the default fallback) and gets a second attempt.

Make's own advice is to pick a fallback model that differs from your primary, so a model-specific outage does not take out both paths (community.make.com, retrieved 2026-10-13). Go further than that and pick a different provider too. If your primary is Claude on your own key, make the fallback an OpenAI model on Make's AI Provider. An Anthropic-wide outage then cannot take out both paths.

The candid con is in the word "once." There is one retry. If the fallback also fails, the run fails exactly as it did in September, so keep the error handlers you already have on any agent that writes to a CRM or sends email. Make's release note also does not say how the failed first attempt meters credits, so check the run history the first time a fallback fires.

Prompt caching: who gets the knob

An agent working through a task makes several model calls, and each one resends the same system instructions and tool definitions. Prompt caching lets Make reuse that stable block instead of reprocessing it, which Make says cuts latency and credit usage (help.make.com, Oct 9, retrieved 2026-10-13).

How much control you get depends on the provider. On supported Anthropic Claude models you choose the cache duration: Off, 5 minutes, or 1 hour (community.make.com, retrieved 2026-10-13). Pass a conversation ID and the cached context carries across repeated calls in that conversation. On supported OpenAI models caching is automatic, with nothing to configure.

For a team standardizing on one provider, that asymmetry matters. Claude users can match the cache to the workload. An agent triggered every few minutes by a webhook keeps a 1-hour cache warm; an agent that runs once a night will not see a cache hit from the night before under either setting. OpenAI users get caching without effort and without a dial. Neither path changes the con the September review raised about premium models: caching trims the repeated input, but a step on Claude Fable 5.1 still draws far more credits than the same step on a cheap model.

The governance half: SSO routing, audit-log MCP tools, event subscriptions

SSO auto-routing closes a narrow but real side door. When an org has a verified domain claim and SSO configured, an employee who clicks Google, Apple, GitHub, or Facebook sign-in with a company email lands in the org's SSO flow with the email already filled in (help.make.com, Oct 9, retrieved 2026-10-13). Make lists it on all plans.

The MCP audit-log tools let admins list audit events, filter them by user, action, or date, and pull the details of one entry, all read-only, from whatever MCP-aware agent or app they already use (help.make.com, Oct 9, retrieved 2026-10-13). A security lead can ask an agent "who changed the billing-sync scenario last Tuesday" instead of exporting logs by hand.

Read the plan line carefully. The release note says the MCP tools are available on all plans. The audit logs they query are Enterprise-only, visible to org owners, org admins, and team admins, and kept for 12 months (help.make.com/audit-logs, retrieved 2026-10-13). On Core, Pro, or Teams there is no audit log to query, so the governance headline applies to Enterprise buyers.

Event subscriptions, from October 8, are the Teams-tier governance item. Make sends signed, real-time notifications to a URL you choose when a subscribed event happens (a scenario is created, a run fails), with a delivery log of every attempt, its status, response code, and payload (help.make.com, Oct 8, retrieved 2026-10-13). Teams gets up to three subscriptions per organization; Enterprise gets unlimited. For a Teams buyer, a run-failed subscription piped to Slack or a pager is the practical alert layer this month adds.

The new models, and one plan wrinkle

Claude Sonnet 5.5 arrived with a context window of up to 1,000,000 tokens and text output of up to 128,000 tokens, in the Make AI Toolkit, Anthropic Claude, and Make AI Agent (New) apps (help.make.com, Oct 9, retrieved 2026-10-13). GPT-6.1 Sol supports Functions, Web search, File search, and Computer use, with multi-agent subtask coordination in beta, in the Make AI Toolkit, OpenAI, and Make AI Agent (New) apps on all plans, per the same release note. The same release also lets Maia review a weak agent run, the tools it used, and its response, then propose a versioned improvement you choose whether to apply.

The wrinkle sits three days later. The October 12 app roundup, adding GPT-6.1 Sol to the OpenAI Create a Prompt module, says it is "available only on paid Make plans," and notes that Free-plan scenarios calling a premium Claude model such as Opus now run on Claude Haiku 5.5 instead (help.make.com, Oct 12, retrieved 2026-10-13). If you are prototyping on Free, the model you pick in the dropdown may not be the model that answers. Test agent quality on Core before you judge it.

Still true from September

The Cycles per run setting was removed on October 1 as announced (help.make.com/2026, retrieved 2026-10-13). If you did not baseline your looping scenarios in September, compare a run from late September against one from this week in the run history and look for a jump in credits per run.

Pricing, retrieved today

The ladder is unchanged since September. Prices are per month for the base 10,000-credit tier; the slider on the pricing page raises credit volume at adjusted rates (make.com/en/pricing, retrieved 2026-10-13).

Plan	Price / month	Credits / month	October-relevant extras	Source
Free	0 USD	1,000	Fallbacks, caching, SSO routing	make.com/en/pricing, 2026-10-13
Core	9 USD	10,000	Same, plus GPT-6.1 Sol in the OpenAI app	make.com/en/pricing, 2026-10-13
Pro	16 USD	10,000	Same	make.com/en/pricing, 2026-10-13
Teams	29 USD	10,000	Up to 3 event subscriptions	make.com/en/pricing; help.make.com, Oct 8
Enterprise	Custom	Custom	Audit logs, so the MCP audit tools work; unlimited event subscriptions	make.com/en/pricing; help.make.com/audit-logs

Where October leaves Make against n8n

On agent reliability, October closes a gap in Make itself: before October 9, a provider outage or an expired key failed the run with no automatic second path. The fallback and the cache now ship inside the managed agent module on every plan, with no extra infrastructure to run (help.make.com, Oct 9, retrieved 2026-10-13). On Make, agent resilience is a setting on a service someone else keeps up. On your own n8n box, keeping the agent and its model connections up is your team's job.

On governance, Make pulls clearly ahead only for Enterprise buyers, because the audit trail is Enterprise-only. Teams buyers get event subscriptions and SSO routing, both useful, neither a full audit story.

On cost, nothing moved. Make still meters every module action as a credit (make.com/en/pricing, retrieved 2026-10-13), and there is still no self-hosted Make. That gap keeps the rating at 4.5 instead of 4.7.

The verdict, by who you are

Make earns 4.5 in October 2026. The picks, by situation:

• Solo non-developer operator. Make. Build on Free, then move to Core at 9 USD a month before you judge agent quality, since Free swaps premium Claude calls to Haiku 5.5 (help.make.com, Oct 12, retrieved 2026-10-13). Set a fallback connection on any agent you leave unattended.
• Mixed-technical ops team. Make, by a wider margin than in September. Teams at 29 USD a month (make.com/en/pricing, retrieved 2026-10-13) now adds run-failed alerts through event subscriptions on top of fallbacks and caching, with no server to patch. Where it flips: if every member is technical and zero per-run cost beats managed governance, self-hosted n8n wins. If you land on Make, Make's paid plans are the buy.
• Enterprise security or IT team. Shortlist Make. Read-only MCP access to a 12-month audit log (help.make.com/audit-logs, retrieved 2026-10-13), plus SSO routing that blocks shadow accounts, gives a security reviewer concrete answers before renewal. Where it flips: a mandate that the workflow engine run on your own infrastructure, which Make cannot meet.
• Technical team that can self-host. n8n still wins on cost, and October does not change that. Once monthly operations clear roughly 30 to 40k and someone can run a Docker box, self-hosting deletes the meter Make charges on. The cost that math leaves out is ops time; if nobody wants the pager, Make's managed agent, now with fallbacks, is still a rational buy.

FAQ

Did Make's rating change in October? No, it holds at 4.5. Fallback connections and prompt caching make Make AI Agents more reliable and cheaper to run on repeat calls, but the two structural cons (no self-hosting, a per-credit meter at volume) are untouched.

How do Make AI Agent fallback connections work? You set a backup connection in the Make AI Agent module's advanced settings. If the primary provider or model fails, Make retries the run once on the fallback, which defaults to Make's AI Provider (help.make.com, Oct 9, retrieved 2026-10-13). If the retry also fails, the run fails.

Does prompt caching work with every model? It works on supported Anthropic Claude and OpenAI models. Claude lets you choose Off, 5 minutes, or 1 hour; OpenAI caching is automatic with no setting (community.make.com, retrieved 2026-10-13).

Can I use the MCP audit-log tools on the Teams plan? The tools are listed for all plans, but audit logs themselves are Enterprise-only (help.make.com/audit-logs, retrieved 2026-10-13), so in practice you need Enterprise to get data back. Teams gets event subscriptions instead, up to three per organization.

What does Make cost in October 2026? Make's Core plan is 9 USD a month for 10,000 credits (make.com/en/pricing, retrieved 2026-10-13), and every October agent feature works on it. Self-hosted n8n can cost less at high volume if someone on the team runs the server.

Review history

This review is updated in place as the product changes. Earlier versions:

• October 13, 2026: AI Agent Fallbacks, Prompt Caching, and MCP Audit Logs (current version)
• September 10, 2026: The ChatGPT Plugin, Claude Fable 5.1, and What October 1 Changes
• August 21, 2026: Maia Builds Your Scenarios, Gemini 3.7 Flash Lands, and Whether the 4.5 Holds
• July 24, 2026: Private Spaces, GPT-5.6, and Whether the 4.4 Holds
• July 17, 2026: AI Agents Go GA and the Rating Moves to 4.4
• June 25, 2026: AI Agents Are Live, Does the Rating Change?