Inception Mercury API Shutdown: December 28, 2026 Wind-Down

Inception told Mercury API customers the hosted diffusion LLM service winds down after December 28, 2026 — while third-party routers already removed Mercury ids and a Vercel AI SDK provider pull request closed the same day citing the shutdown. The news spread through Hacker News on November 12, 2026 before Inception published a standalone blog explainer.

TL;DR

What customers reported

HN user _matthew_ posted the full email text attributed to hello@inceptionlabs.ai:

• Mercury API remains available through December 28, 2026
• After that date, customers lose service access and data submitted to the platform
• Inception encourages exporting necessary data before the cutoff

A shorter Discord message from Yannis Miraoui quoted on HN adds that OpenRouter access ended November 8, 2026 for Mercury models — aligning with third-party catalog trackers that saw routes disappear after listed expiry dates.

No matching press release was indexed on Inception's models page at publication time — the marketing site still describes Mercury 3, Decide, and Voice, which creates claimed vs verified tension for new signups.

What developers are saying on Hacker News

These are opinions from the November 12 thread, not verified facts:

• johnshahbazian wrote that they rely on Decide and wondered whether Inception is being acquired.
• stilesja reported heavy Mercury 3 use and speculated about an acquisition after the abrupt email.
• wing_rets said Mercury was cheap and fast and asked whether the tech might be open sourced.
• carterschonwald praised Mercury's coding style (admitting gaps, looking things up) and hoped for an IP release.

explainx.ai treats acquisition and open-weight rumors as unconfirmed until Inception posts primary documentation.

Ecosystem impact the same afternoon

Vercel contributor aayush-kapoor opened a draft @ai-sdk/inception provider on November 12, then closed PR #24975 with comment "Mercury API is winding down - closing this PR" — a fast signal that framework maintainers will not ship first-party Inception support while the API sunsets.

Teams that pinned Mercury for diffusion decoding speed in agent loops should assume no new SDK surface area from Vercel this quarter.

Migration checklist

1. Inventory calls — grep for inceptionlabs.ai, mercury-3, and OpenRouter slugs.
2. Export datasets — prompts, eval transcripts, and billing CSV before December 28.
3. Rerun latency baselines — diffusion LLMs marketed 1,000+ tokens/s (Mercury 2 coverage); replacements may be autoregressive.
4. Voice stacks — if you used Mercury Voice, map time-to-first-token requirements to alternate duplex vendors such as Fish Audio Live 3.
5. Decision APIs — Mercury Decide workflows need explicit confirmation whether Decide endpoints sunset on the same date.
6. Legal — enterprise agreements may differ from self-serve email terms quoted on HN.

Replacement model strategies

None are mechanical drop-ins; budget two weeks of eval before December.

Business context

Inception marketed Mercury 3 in October 2026 with 50% launch discounts and 1M context claims we analyzed in claimed vs measured coverage. A shutdown email six weeks later is a reminder that speed-first API startups can retract hosting even while docs pages still upsell.

Deprecation sunset graphic for hosted diffusion LLM API wind-down

What people are asking (November 12)

What remains unverified

• Whether Mercury Decide, Voice, or enterprise contracts follow the same sunset
• Any acquisition buyer or open-weight release
• Official Inception blog or status page mirroring the HN-quoted email
• Whether self-hosted or partner-hosted Mercury inference will appear elsewhere

We will update this post when Inception docs publish a shutdown notice matching the email.

Timeline leading to November 12

The sequence looks like distribution choke first (OpenRouter), hosted API second (December), with marketing pages lagging reality.

What diffusion LLMs changed in agent loops

Mercury pitched parallel token decoding unlike classic autoregressive chat models. Teams adopted Mercury for:

• Tight agent loops where milliseconds per step compound
• Voice stacks needing low time-to-first-token (Live 3 contrast)
• Cheap bulk classification adjacent to Mercury Decide

If you must migrate to autoregressive APIs, expect higher tail latency unless you adopt speculative decoding elsewhere — see NVIDIA Nemotron diffusion research for how labs still experiment, without guaranteeing hosted APIs.

Mercury Decide and Voice — separate products?

Inception docs still list Mercury Decide and Mercury Voice with pricing tables on November 12. The HN email refers to Mercury API generically. Enterprise buyers should open tickets asking:

1. Do Decide endpoints sunset on the same calendar date?
2. Are Voice enterprise contracts carved out?
3. Will hosted inference move to a partner domain?

Until answers land, treat all Inception hosted inference as at-risk.

Data export and compliance

The quoted email says customers lose data submitted to the platform after December 28. That may include:

• Prompt/response logs used for evals
• Fine-tuning datasets uploaded for experiments
• API usage records needed for finance audits

Export early; do not rely on a last-week scrape. If you operate under GDPR or HIPAA, align with counsel on whether shutdown notices satisfy processor change requirements.

Rewriting routers and SDK config

Search your monorepo for:

• api.inceptionlabs.ai
• INCEPTION_API_KEY
• mercury-3, mercury-2, mercury-voice
• LiteLLM provider blocks naming Inception

Update fallback chains so production does not call a dead host at year-end traffic spikes. Stage drills before Thanksgiving week if your team freezes deploys in late December.

Financial and vendor-risk lessons

• Promotional pricing on young APIs is not a multi-year contract.
• Speed benchmarks in marketing do not imply business continuity.
• Third-party routers can remove models before the vendor email arrives — design routers to fail loudly, not fall back to mystery defaults.

Monitoring and alerting before the cutoff

Add synthetic checks that call a read-only Mercury endpoint weekly until December 20, then daily. Alert if:

• HTTP 402/403 rates spike
• Latency doubles (early deprecation throttling)
• Error bodies mention scheduled maintenance without a public status page

Pair with Bedrock agent observability patterns if you move inference to AWS-hosted models — different metrics, same need for per-step model id logging.

Contract language to re-read this week

Enterprise agreements sometimes define wind-down notice periods, data retention, and refund clauses separately from self-serve email text. Legal teams should compare:

• Master services agreements signed before Mercury 3 launch
• POC addenda referencing Mercury 2 sunset dates
• Partner contracts naming Inception as a subprocessors

If your contract lacks a successor vendor clause, negotiate transition assistance in writing before December — verbal Discord answers are not procurement evidence.

If Inception open-sources weights later

HN commenters hope for an IP drop. If that happens, self-hosting still requires GPU ops, safety review, and license terms unknown today. Budget engineering time only when a primary Inception or buyer announcement names a repo — not when rumors circulate in shutdown threads.

Usage quota hourglass approaching zero for API credits

Related reading

• Mercury 3 launch claims
• Mercury 2 diffusion speed
• Diffusion LLM primer
• Gemini 4 Hydrogen fast tier
• OpenRouter-style routing economics

Product managers should schedule a thirty-minute migration review with engineering and finance before Black Friday traffic: confirm which customer-facing features still call Mercury, what substitute model fits each SLA, and whether marketing pages must be edited so new signups do not advertise a dead API. Silence from Inception's blog does not relieve teams of customer communication duties once the shutdown email lands in inboxes.

Shutdown dates quoted from customer reports on November 12, 2026 until Inception posts an official mirror.