Keep the gateway, test the model
A shared gateway lets you keep your TokenSense base URL while changing the model behind a compatible request. It also keeps attribution and spend controls in one place. That reduces integration work, but it does not make different models interchangeable.
Start with a copy of the working automation and a set of tasks whose correct outcomes you can check. Test the complete tool loop, structured output and error branch before changing production traffic. A cheaper token rate only helps if the model completes the job reliably; an HTTP 200 alone does not establish that.
An October 2026 text-cost example
This example compares Claude Sonnet 5.5, GPT 6.1 Sol, and Gemini 3.8 Flash using 1,000 uncached input tokens and 250 billable output tokens per task, across 180,000 tasks per month. Output includes any billed thinking tokens. These are standard text rates checked on 4 October 2026, not benchmark scores or a prediction of your bill.
Sonnet 5.5 and GPT 6.1 Sol each use $2 input and $10 output per million tokens in this example: $4.50 per 1,000 tasks, or $810 per month. Gemini 3.8 Flash uses $0.75 input and $3.75 output: about $1.69 per 1,000 tasks, or $303.75 per month. The $506.25 difference assumes identical billable usage; real models can use different amounts.
Google lists those Gemini launch rates through 31 December 2026, with $1.50 input and $7.50 output starting 1 January 2027. Cache reads and writes, long-context tiers, batch or priority service, images, audio, and tool fees are excluded from this estimate.
Compare models for the same token workload
1,000 input + 250 billable output tokens per task
Estimated standard text spendRates checked 4 October 2026: OpenAI, Anthropic, Google. Gemini launch rates expire 31 December 2026. Excludes cache, long-context, service tiers, non-text usage, and tool fees. Output includes billed thinking. GPT 6.1 Sol tools require Responses; this example compares text without tools.
Match the endpoint to the workflow
GPT 6.1 Sol and GPT 6 Astra tool requests require the Responses API. A tool-enabled n8n Chat Model workflow therefore needs a Chat-compatible model, or a separate HTTP Request integration using /v1/responses. Text-only GPT 6.1 Sol requests can use Chat Completions.
Gemini 3.8 Flash supports low, medium, and high thinking. TokenSense maps minimal to low and removes obsolete sampling controls for this model. Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Opus 4.7 fix sampling and count thinking toward the output limit. Inspect the x-tokensense-adapted response header when diagnosing a request adjustment.
Images are a separate operation. Gemini image models need Images or native generateContent, rather than Chat Completions. The TokenSense Images compatibility route currently accepts one square 1K Gemini image. Provider account access and quota still apply.
Compare measured runs, including unknown costs
Use the same test set and project tag for each model. Compare successful task completion as well as latency, usage, and cost. Retest multi-turn tools and fallback routes: model changes can affect reasoning, tool-result formats, and output shape.
Some successful image or audio responses lack usage needed to calculate the full bill. Mini TTS binary audio is one example. Preserve unavailable pricing as unknown rather than converting it to zero. In n8n 0.1.20, image and transcription output exposes pricingAvailable and billingReason; speech output uses the gateway billing headers when present.
Roll out deliberately
Keep an export of the working workflow and change one model at a time. Confirm the selected model is available to your provider account, then run a limited production sample with a budget cap. Review failures and output quality before widening traffic.
Confirmed retired models return model_retired. TokenSense keeps stored model IDs for history rather than silently migrating saved workflows. Select a supported replacement and verify its request format; changing an Imagen ID to a Gemini image ID is not enough.
October model update and n8n release status · n8n node reference
Write down the contract before changing the model
An OpenAI-compatible interface reduces connection work. It does not guarantee equivalent parameters, tools, output or task quality. Write a short contract for the workflow: what it receives, which tools it may call, which fields it must return, what counts as success and what it should do when the model cannot answer.
For example, a support triage agent must retrieve the matching order, return a status from a fixed set, include the verified delivery date or null, and send uncertain cases to a human. Those are testable requirements. “Produces a helpful reply” is too vague to catch a wrong order or an invented date.
Keep a baseline export containing the current node versions, model ID, endpoint, prompts, tool schemas, memory settings, output parser, routing rules and retry settings. When several workflows share credentials or routing rules, record which ones depend on them before changing the shared configuration.
Use this compatibility checklist
Before switching providers, check endpoint and account access, supported request controls, tool arguments and results, multi-turn state, structured output, streaming and error handling, and complete usage accounting. Test these through your actual automation client and gateway, then compare successful task outcomes before rolling out.
1. Endpoint and access: confirm the exact model ID, provider key permissions, region and route. A model listed in a catalogue may still be unavailable to your account. Validate the tool-enabled request through the endpoint the workflow will actually use.
2. Request controls: check reasoning effort, sampling and output limits for that model. Do not assume a setting was honored just because a request succeeded. TokenSense reports adaptations in x-tokensense-adapted; inspect it when diagnosing differences.
3. Tool contract: validate names, argument schemas, required fields and result handling. Test no tool call, one tool call, multiple calls and a tool failure. Preserve the association between a tool result and the call it answers; some models require reasoning state alongside the tool exchange.
4. Multi-turn state: test the second and third turns, not just the first response. Gemini thinking models use thought signatures; the complete client and gateway path must preserve the state required by the model. Keep conversation and tool history in the format that path supports.
5. Output contract: validate the exact JSON fields, types, enums and nullable values your next node consumes. Test an empty answer, missing data, a refusal and an output limit. A prose answer that looks right can still break the next node.
6. Streaming and errors: check whether the client gets complete tool arguments, a complete final answer and a usable failure signal. Exercise a timeout and an unavailable model in a test environment. Confirm which failures are retried and which route to a human.
7. Usage and cost: record billable input, cache categories, output including any billed reasoning, tools and unknown pricing. Compare cost per successful task, not only price per million tokens. Check what remains outside the gateway, such as CRM or search-service charges.
Google documents its own OpenAI compatibility limitations. Treat provider compatibility, gateway translation and your automation client as separate things to verify. A feature supported by the provider may still be missing from the client’s node or the gateway route.
A practical acceptance test for a support workflow
Duplicate a support workflow and replace message sending with a draft output. Point lookups at test records. Use the same tool definitions and downstream parser as production. Then run both models against the same saved inputs; inspect the outcomes yourself instead of asking one model to decide whether its own response is good.
Case A: a customer has one matching order. Pass if the agent fetches that record and returns its actual status and date in the required fields. Fail if it answers from memory or makes up a date.
Case B: two orders match the name. Pass if the workflow asks for an order identifier or routes to a human, according to your contract. Fail if it confidently selects an arbitrary order.
Case C: the order lookup returns no delivery date. Pass if the result contains null and the intended escalation. Fail if the output parser accepts an invented date merely because it is well-formatted.
Case D: the lookup times out after a tool request. Pass if the error path follows the retry policy and does not duplicate an external action. Inspect the later turn as well as the initial tool call.
Case E: repeat a task with a long history and a deliberately restrictive output limit. Confirm that the downstream step handles truncation or missing output. This tests a boundary that short, fresh conversations can hide.
These are proposed test cases, not a claim that any named model passed them. Add cases from your own failures. Run tasks more than once when output varies, and agree acceptable error, latency and cost thresholds before you compare the candidates.
Make the cost comparison about completed work
For your sample, divide all measured model spend, including billable failed and repaired attempts, by completed tasks that meet the contract. Keep unknown-cost attempts visible beside the result. This avoids picking a model that looks cheap per request but repeatedly needs another pass.
An illustrative comparison: model A costs $0.02 per attempt and takes one attempt for each of 100 successful tasks: $2 total, or $0.02 per success. Model B costs $0.012 per attempt, but uses 200 billable attempts to complete those same 100 tasks: $2.40 total, or $0.024 per success. These invented workloads demonstrate arithmetic, not model quality or expected retry rates.
The n8n agent cost walkthrough shows how to sum the model turns inside a run. Use actual usage from each candidate because prompts may tokenize differently, tools may run in a different sequence and reasoning output may differ. The fixed-token comparison above isolates rates; it does not establish a real saving.
Keep fallback and rollback inside the test
Fallback needs the same compatibility checks as a planned switch. A second provider must accept the workflow’s tools and return an output the next node can use. Define when fallback is allowed, how many attempts it may make and whether the original attempt could already have completed an external action.
Distinguish a model failure from a local policy block. A budget denial should follow your stop or escalation path; changing provider should not bypass that budget. Follow the error reference for the status and retryability your node exposes rather than treating every error as “try another model.”
TokenSense error codes and retry handling
When the candidate passes your checks, move a limited, reversible portion of work first. In a no-code setup, that can mean one duplicate workflow with one test client or internal queue. Keep the old workflow export and credentials available. If correctness, latency, unknown usage or repair rates cross your agreed threshold, restore the prior configuration and inspect the failing executions before widening the rollout.
Try one compatible workflow through TokenSense
TokenSense provides a shared gateway with attribution and budget controls for routed requests. It simplifies the connection and accounting side of a switch; your acceptance tests still establish whether the replacement completes the job. Start with one copied workflow and compare its attributed requests and actual outcomes.
Start free with TokenSense to inspect routed usage, then follow the n8n setup guide or the Make and Zapier guide for the connection your workflow uses. Project and key budgets require Pro or Agency.

