Topic
OpenAI-compatible gateway products that unify provider access, failover, budgets and observability for agent traffic.
Vercel added five named Microsoft MAI models to AI Gateway on October 1, 2026, covering speech, transcription and image generation. Each modality bills in a different unit, so teams have to estimate cost feature by feature.
Reality
Vercel offers the Glyph Cluster stealth model free on AI Gateway, but prompts and responses may be used for training and ZDR is unavailable. Any trial is a data-governance decision first, so one dev.to plan limits testing to public, non-sensitive material.
Reality
OpenAI's Decisions API answers yes/no, pick-one and scale questions for $0.10 per million input tokens, with output free. Jev sells the same three question types for text at less than half that, so OpenAI's case rests on image input and compliance terms.
Perspective Coverage
4 publishersReality
Vercel's AI Gateway now reruns successful but low-confidence decisions on a second model and bills both runs. Each escalation adds a billed call and its latency, so the threshold belongs on decisions where labelled tests show the second run fixes costly errors.
Reality
Cloudflare now sells web search from Ceramic.ai, Exa and Linkup through AI Gateway at partner list prices, billed from gateway credits with no markup. Teams already on the gateway skip a separate search vendor account, though their agents still have to issue each search call.
Reality
Cloudflare's Auto Router, now in public beta, picks a model per AI Gateway request and cut costs by up to 30% in the company's own OpenCode use. How much of that transfers depends on how much of a team's traffic never needed a frontier model.
Reality
LiteLLM's MCP test endpoints let any valid proxy key run arbitrary commands on the gateway, rated CVSS 8.8. CISA added the flaw to its Known Exploited Vulnerabilities catalog on June 8, 2026, confirming exploitation in the wild.
Reality
Wiz found 294 of 3,074 internet-facing LiteLLM gateways answering admin requests in February, and two thirds of those had no key set at all rather than the documented sk-1234. That changes what needs fixing.
Reality
OpenAI's guide names Sunburst for editing precision and Flare for speed, and both run from either API. The editing loop, File ID inputs and the action switch exist on the Responses side alone.
Perspective Coverage
7 publishersReality
MetTel CTO Ed Fox says AI agents need scoped permissions and audited trajectories, citing sandbox lapses involving Anthropic, OpenAI and Moonshot AI models. His privileged-user model handles access granted by mistake, but agents that try another route when blocked need monitoring built for that behavior.
Reality
LiteLLM works as a drop-in OpenAI replacement for teams running their own clusters, while managed gateways suit teams renting inference. OpenRouter's reported $113M round at a $1.3B valuation funds the rented side.
Reality
Fastly's AI Runtime Control is an endpoint you point an application at. It issues virtual keys against your own model provider account, and a key can carry a dollar budget, a rate limit and an ordered list of fallback models.
Reality
AI Gateway Budgets went generally available with ceilings at team, project, key and teammate scope. The check runs at the start of each request, so a tight loop can finish slightly past the number you set.
Reality
A dev.to post argues MCP's attack surface follows from point-to-point model-to-server links and belongs behind an inline gateway. The tool-poisoning path it documents runs through description text that a gateway may forward unchanged.
Reality
A dev.to post credits a roughly 42x drop in token use on a code-editing workload to deleting an agent's reread loop. The gateway it recommends can only route whole calls, and the two levers are different sizes.
Reality
Enterprise AI pricing has separated the seat from the usage, which moves the cost driver from headcount to demand. The overruns now on the record suggest buyers tend to find out months after the fact.
Reality
Centralising prompt checks and PII filters at the gateway mirrors putting TLS at the edge, and the arithmetic backs that comparison. The semantic checks themselves, though, work nothing like TLS termination.
Reality
The router picks a cheaper model for work that resembles work it has already seen, and falls back to the strongest one when nothing matches. The same traces that enable that call are the proposed training set.
Reality