Fireworks Forge: Faster Agents, Half-Price Background Runs

Fireworks used its first Forge conference on November 3, 2026 to attack the two costs that hurt agent builders most: waiting and paying full price for work that can wait. The headline launches are Ultra-fast inference (public preview), a half-price Background Mode, multi-day prompt caching, and Nexus Desktop, which routes open and closed models inside the Codex app, Claude Desktop and OpenCode. If you run multi-step agents on open-weight models, this changes how you should split your traffic.

What changed

• Ultra-fast (public preview): inference on NVIDIA Blackwell GPUs that Fireworks says reaches up to 1,500 tokens per second on DeepSeek V4.1 Flash.
• Background Mode: a lower-cost runtime that Fireworks describes as half price, for work that can wait, such as eval data generation and overnight research.
• Multi-day prompt caching and higher serverless rate limits (the rate limits are generally available) for agents that pause and resume.
• Elastic Burst (generally available): usage can exceed its normal ceiling when spare capacity exists, with no code or pricing changes.
• Nexus Desktop: joins the Nexus CLI and plugs model routing into the Codex app, Claude Desktop and OpenCode. Pro plans start at $8 a month, Teams at $10 per seat per month.
• Specialized Model Inference (public preview): serverless serving of trained LoRA adapters with per-token pricing and no charge while idle, starting with GLM 5.3 Flash.
• No Eval and Fireworks Sovereign (both private preview): evals generated from production traces, and the Fireworks inference engine running in your own cloud or data center.

Why latency is the real agent tax

A chat app makes one model call per answer. An agent makes dozens: plan, call a tool, read the result, retry, summarize. Every call adds its generation time to the total, so a model that feels fast in a chat window can feel slow inside a 30-step loop. That is the pitch behind Ultra-fast. Fireworks says a benchmarked 30-turn agent task fell from about one minute to eight seconds. That is the vendor's own benchmark on its own stack, and the post does not publish the task or the baseline configuration, so treat it as a direction rather than a number to plan capacity around.

The practical point still holds. If your agent's wall-clock time is dominated by output tokens, throughput matters more than a small quality gap between models. Our guide to open-weight coding models covers which models are worth putting in that loop.

Background Mode: split your traffic by urgency

The more useful launch for most teams may be the cheap one. Plenty of agent work has no human waiting on it: generating synthetic eval sets, nightly codebase audits, bulk summarization, research jobs that report in the morning. Fireworks prices that work at half its normal rate in Background Mode. The announcement does not list which models are covered or give per-token rates, so check the pricing page before you assume a specific saving.

The pattern this enables is simple: tag each job as interactive or deferrable, and route them to different tiers. An illustrative sketch (not tested; the routing field is a placeholder, so check Fireworks' docs for the real parameter):

if job.user_waiting: tier = "ultra-fast" # latency first

elif job.deadline_hours >= 8: tier = "background" # half-price tier

else: tier = "standard"

Combined with multi-day prompt caching, which avoids paying again to reprocess a long context each time a paused agent resumes, this targets the two biggest line items in a long-running agent bill.

Nexus: model routing inside the tools you already use

Nexus is Fireworks' answer to a different problem: developers want to use cheaper open models without leaving Codex, Claude Desktop or OpenCode. Nexus sits between the tool and the model, chooses a model per task, and gives the organization one place to see spend by model. Fireworks says it compares that spend against an estimated all-Opus 5.5 baseline, which is a useful framing for finance but is still an estimate, not an invoice.

At $8 a month for Pro, the subscription itself is cheap; the real cost is the model usage it routes. If you already use a gateway such as OpenRouter or Vercel AI Gateway for apps, Nexus is aimed at a different layer: the individual developer's desktop and CLI tools rather than your production API traffic. Our AI gateways comparison covers the app side, and our roundup of AI CLI coding agents covers the tools Nexus plugs into.

How it compares

• Against closed-model APIs: Fireworks' bet is that open models plus routing and training beat paying frontier prices for every call. That works when your tasks are well defined enough to evaluate.
• Against gateways: OpenRouter and Vercel AI Gateway route across many providers; Fireworks routes across models it serves itself, with its own speed tiers. You can use both.
• Against self-hosting: Sovereign is Fireworks' move into your own infrastructure, but it is a private preview with no public pricing, so it is not yet a plan you can budget.

Pros and cons

• Pro: a clear speed tier and a clear cheap tier let you price agent work by urgency.
• Pro: idle LoRA adapters cost nothing to keep served, which lowers the cost of experimenting with fine-tunes.
• Pro: free training on Muse Glimmer 30B through November 30, 2026, per the announcement.
• Con: the speed and task-time numbers are Fireworks' own benchmarks with limited methodology.
• Con: Ultra-fast is a preview with no per-token price in the announcement; No Eval and Sovereign are private previews.
• Con: Background Mode model coverage is described only as eligible work.

Verdict

If your agents run on open-weight models, Forge is worth acting on now in one specific way: separate your deferrable jobs and test Background Mode on them, since a half-price tier for work nobody waits on is the lowest-risk saving here. Treat Ultra-fast as something to benchmark on your own loop before committing, because the headline numbers are vendor-run. Nexus is worth a trial for teams whose developers already juggle Codex, Claude Desktop and OpenCode and want one view of model spend. See the Fireworks AI company page for background.

Sources

• Fireworks: Fireworks unveils new products at Forge (Nov 3, 2026)
• Fireworks on X: Forge announcement
• The Agentic Weekly: Fireworks Forge writeup