Fireworks Forge: 1,500 Tokens/Second Inference and Half-Price Agents

Fireworks AI has a specific customer in mind: the team running dozens of model calls inside every agent task, watching the token bill climb and the latency stretch. At Forge, its first annual conference on November 3, 2026, the company attacked both problems at once: a speed tier that pushes up to 1,500 tokens per second, a cheap tier that runs deferrable work at half price, a router that brings open models into the tools developers already use, and an option to run the whole stack on your own GPUs. Here is every launch, what each one actually costs, and where the fine print lives.

What Forge announced

The lineup splits into four themes: faster inference, cheaper inference, model routing for everyday work, and training plus enterprise deployment. Each item's availability state is as Fireworks described it:

• Ultra-fast inference, public preview, on NVIDIA Blackwell GPUs
• Background Mode, a half-price runtime for work that can wait
• Multi-day prompt caching and higher serverless rate limits, generally available
• Elastic Burst, generally available
• Nexus Desktop, new, joining the Nexus CLI with Pro plans from $8/month
• Ember-1, Fireworks' own model on serverless and dedicated deployments
• No Eval, private preview
• Specialized Model Inference, public preview, starting with GLM 5.3 Flash
• Model governance, spending controls and exportable audit logs, generally available
• Fireworks Sovereign, private preview

Previews can change without much warning, so treat every preview feature as a moving target until it reaches general availability.

Ultra-fast: the speed tier

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 Ultra-fast reaches up to 1,500 tokens per second as demonstrated on DeepSeek V4.1 Flash, and that completion time for a benchmarked 30-turn agent task fell from one minute to eight seconds. Both numbers are the company's own, from its own stack, with the task and baseline configuration unpublished, so treat them as a direction rather than numbers to plan capacity around. (For context on that model, see our DeepSeek V4.1 Flash 1017 review.)

The pricing side is where the launch gets expensive. Fireworks put no per-token rates for Ultra-fast in the announcement. Independent analysis by S3DAT of Fireworks' published pricing table found that Ultra-fast on DeepSeek V4.1 Flash 1017 bills at exactly eight times the standard rates. The published standard rates for the 1017 model are $0.30 per million uncached input tokens, $1.20 per million output tokens, and $0.006 per million cached input tokens, which puts Ultra-fast at $2.40, $9.60, and $0.048 per million respectively.

For a hypothetical workload of one million uncached input tokens plus one million output tokens, that is $12 instead of $1.50: an extra $10.50 to buy the faster completion, holding token counts fixed. Fireworks positions Ultra-fast for latency-sensitive work, says output quality remains the same, and targets generation speeds six to eight times higher. Note that generation speed measures time spent producing tokens, not an entire workflow that also waits on tools and external services. Ultra-fast sits in Fireworks' Medium rate-limit tier (ceilings of 43.2M prompt, 10.8M uncached, and 432K generated tokens per minute at Tier 3 and above, per S3DAT), and the pricing table lists no Priority option for it.

Background Mode: the cheap tier

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. Background Mode is a lower-cost runtime Fireworks describes as half price compared with its normal rate, aimed at exactly this deferrable work.

The announcement does not list which models Background Mode covers or give per-token rates, so check the pricing page before assuming a specific saving. "Half price" is relative to standard rates, not a published table of its own.

Two cost-cutters pair with it. Multi-day prompt caching lowers costs for repeated context, since persistent agents pay to reprocess their context every time they resume after a pause. Unite.AI flagged a Fireworks benchmark showing that in one DeepSWE run, 99.6% of input tokens were cache hits and cache-hit charges were 60% of the bill, which makes repeated context a material line item for long-running agents. Higher serverless rate limits, now generally available, let teams run more work concurrently without reserving dedicated capacity, and Elastic Burst, also generally available, allows usage to exceed the normal ceiling when spare capacity exists, with no code or pricing changes.

The pattern this enables is simple: tag each job as interactive or deferrable, and route them to different tiers. Latency-first jobs go to Ultra-fast, overnight jobs go to Background Mode, and everything else stays on standard.

Nexus Desktop: routing where you already work

Nexus is Fireworks' answer to a different problem: developers want to use cheaper open models without leaving the tools they already use. It combines open and closed models with personalized model selection and cache awareness, and gives organizations centralized spending visibility. Fireworks shows spend by model compared against an estimated all-Opus 5.5 baseline, which is useful framing for finance but still an estimate, not an invoice.

Announced in July, Nexus previously lived in the CLI. Forge adds Nexus Desktop, which runs model routing inside the Codex app, Claude Desktop and OpenCode. Per the Nexus Desktop documentation, connections include ChatGPT, Claude Code, Cursor IDE, VS Code and OpenCode, with Claude Desktop support described as limited. The desktop app ships as a universal macOS build for Apple silicon and Intel Macs, with Linux and Windows support coming soon. The CLI runs on macOS and Linux, with limited Windows support through Git Bash.

Pricing is the new part: Nexus Pro plans start at $8 per month (about £5.93 or C$11.16 at November 4, 2026 market rates), Teams plans start at $10 per seat per month (about £7.41 or C$13.95), and Enterprise plans add single sign-on, governance, shared budgets and access controls. The critical fine print is that the subscription covers routing and visibility, while the model usage it routes still bills separately. At $8, the subscription itself is cheap. The real cost is the usage behind it.

OpenRouter and similar services route across many providers' models; Fireworks routes across models it serves itself, with its own speed tiers. You can use both. Nexus is aimed at the individual developer's desktop and CLI tools, while gateways like the ones in our KT Token Factory coverage sit in front of production API traffic.

Train on your own data

The second half of Forge is about making models better at company-specific tasks rather than just serving them faster. No Eval, now in private preview, turns production traces and product signals into task-specific evaluations, a direct answer to the complaint that writing your own evals can take weeks of data labeling and grading.

Specialized Model Inference, now in public preview, extends a serverless approach to serving LoRA adapters, starting with GLM 5.3 Flash. Teams evaluate a trained model and serve it through the Fireworks Serverless API with per-token pricing at the base model's serverless rates, no dedicated deployment to size, and no serving charge while it sits idle. The documented rates for GLM 5.3 Flash are $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens (about £0.11/£0.02/£0.37 or C$0.21/C$0.04/C$0.70). The preview supports a maximum LoRA rank of 64 and up to three loaded adapters per account, with loading in under five minutes, though there are no latency guarantees during the preview and the first request on a fresh adapter can take up to a minute.

To get teams started, Fireworks is offering free training on Muse Glimmer 30B through November 30, 2026, including supervised fine-tuning, direct preference optimization and reinforcement learning. That is a genuine free tier with a hard expiry date: use it or lose it.

Fireworks also talked up Ember-1, its own model on serverless and dedicated deployments, which it says used about 40% fewer tokens than Kimi K3 at comparable quality. That is Fireworks' own benchmark, not an independent reproduction. (Kimi context in our Kimi K3.1 Pro review.)

Sovereign: Fireworks on your GPUs

For enterprises that want to run on their own infrastructure, Fireworks Sovereign, now in private preview, brings the company's inference engine to the customer's cloud, neocloud, or data centers. Inference and training stay inside the customer's environment under its existing security and change-control policies, which is the whole pitch for regulated industries with data-residency requirements. Fireworks says it already manages more than 25,000 GPUs in customer environments.

The company claims Sovereign delivers 20 to 50% more throughput on identical hardware and about 2.3 times the tokens per hardware dollar. Again, those are Fireworks' own figures, with no independent verification available yet. There is no public pricing for Sovereign at all, so it is not something you can budget for today. It is the strategic long bet in the lineup, while the near-term value sits in the serverless tiers.

The price breakdown

Currency conversions use November 4, 2026 market rates, verified at publication (1 USD = £0.741 = C$1.395). Your provider's actual rate will differ.

How to use it

If you run agent workloads on open-weight models, the lowest-risk move is the cheap one: separate your deferrable jobs, eval-data generation, overnight research and batch audits, and test Background Mode on them before committing interactive traffic to anything new. For Ultra-fast, applications select its dedicated router name in the model field. Benchmark it on your own loop rather than trusting the vendor's 30-turn task, especially if your workflow spends most of its time waiting on tools rather than generating tokens.

Teams whose developers already juggle the Codex app, Claude Desktop and OpenCode can trial Nexus Desktop to get one view of model spend, keeping in mind the $8 and $10 subscriptions route usage but do not include it. If fine-tuning is on your roadmap, the free Muse Glimmer 30B training through November 30 is the time-boxed window, and the docs are the place to confirm exact parameters before you start a run. For a comparable MCP-based tooling story on the Google side, see our Google Data Agent Kit guide.

What the announcement does not answer

Honest gaps worth knowing before you plan around Forge. The Ultra-fast numbers are demonstrated on one model, DeepSeek V4.1 Flash 1017, and coverage of other models is unspecified. Background Mode's eligible models are not listed, and its per-token rates are described only as "half price" relative to standard. There is no Priority option for Ultra-fast. Serverless LoRA adapters carry no latency guarantees during the preview. Benchmark methodologies, the 1,500 tokens per second figure, the 30-turn task, and Ember-1's comparisons, are all Fireworks' own with no independent reproduction published. And Sovereign has no public pricing at all. None of this invalidates the launches; it sets the boundary between what is proven and what is promised.

Common mistakes

• Reading the $8 Nexus plan as all-inclusive. It covers routing and visibility. Your model usage bills on top, and that is where the money goes.
• Assuming half price applies to everything. Background Mode covers "eligible work." Confirm your model and job type before budgeting the saving.
• Benchmarking Ultra-fast on a tool-bound workflow. If your agent spends most of its time waiting on API calls and databases, faster token generation barely moves the wall clock. The win shows up when generation dominates.
• Treating 1,500 tokens per second as a workflow speed. It is a generation-speed figure on one model in a vendor benchmark. Measure your own loop before committing.

What is Fireworks Ultra-fast?

A public-preview inference tier on NVIDIA Blackwell GPUs that Fireworks says reaches up to 1,500 tokens per second on DeepSeek V4.1 Flash. It bills at eight times standard token prices, $2.40/$9.60/$0.048 per million input/output/cached-input tokens on that model, with no Priority option and Medium rate-limit tiering.

How much does Fireworks Nexus cost?

Nexus Pro plans start at $8 per month and Teams plans at $10 per seat per month, per the November 3, 2026 announcement, with Enterprise plans adding SSO, governance, shared budgets and access controls. Model usage is billed separately from the subscription.

What is Background Mode?

A lower-cost Fireworks runtime for AI work that can wait, such as eval-data generation or overnight research. Fireworks describes it as half price compared with its normal rate. The announcement does not list which models it covers.

Is Ultra-fast worth the 8x price?

It depends on your workload. If your agents are generation-bound and humans are waiting, six-to-eight-times faster generation can justify the premium. If your loop is dominated by tool calls and external waits, it mostly will not. Run your own loop before committing, since the headline numbers are vendor-run.

How do I try the Forge launches?

Check Fireworks' pricing page first, select the Ultra-fast router name in the model field for latency-first jobs, and use the Nexus Desktop macOS build to route across the tools you already use. The free Muse Glimmer 30B training runs through November 30, 2026.

Forge is Fireworks' bid to make inference economics a first-class decision: match the speed tier and the price tier to each workload instead of paying one flat rate for everything. The launches with published prices, Nexus subscriptions, GLM 5.3 Flash LoRA serving, and the 8x Ultra-fast math, are the ones you can act on today. The previews are worth watching, not budgeting around. If your agents run on open models, start with the cheap tier.

Sources

• Fireworks, "Fireworks unveils new products at Forge" (official announcement, November 3, 2026) — fireworks.ai
• Metarix, "Fireworks Forge: Ultra-fast and Background Mode" (November 4, 2026) — metarix.io
• S3DAT News, "Fireworks opens Ultra-fast preview at eight times standard token prices" (November 4, 2026) — news.s3dat.com
• Unite.AI, "Fireworks Announces Faster Agent Inference and Sovereign Deployment" (November 3, 2026) — unite.ai
• MarkTechPost, "Fireworks AI Releases Fireworks Nexus" (July 28, 2026) — marktechpost.com