NEWMEN

Proof

Proof, not promises.

Every claim points at a receipt you can see. We pointed real coding agents at Newmen and built the same apps twice — once going direct, once through Atlas mode. Below: the head-to-head case study, illustrative cost benchmarks, a value and speed leaderboard, and the latency & reliability story. You pay less than going direct from your first call, you see the model that served it, and if you don’t like an answer it’s free. Think Cloudflare for AI inference.

The guarantee

Two guarantees, in writing.

Every call is cheaper than going direct and you see the exact savings — thumbs-down anything you don't like and it's free, no questions.

Price

Never more than going direct — and you start saving on call number one. Atlas takes at least 5% off from the first request and climbs as it ramps. The savings show up line-by-line in your console.

Quality

Thumbs-down any call you don't like and we refund it in full — no ticket, no explanation, no argument. We guarantee quality instead of asserting it.

Migration

We migrate you in under a day, or we run free until you're live. Two env vars, your existing client, every call logged with what it cost versus direct.

  • Never more than going direct
  • Thumbs-down → full refund
  • No card to start · no subscription

Case study

Claude Code. The same app, built twice.

We built a genuine app with Claude Code once going direct and once through Atlas mode — same prompt, same client. Only the bill changed, and every figure is pulled straight from the console's own usage records, trade-offs and all. More agents are in the works.

Newmen · Atlas case study

Same app.Smaller bill.

We pointed Claude Code at Newmen with nothing but an API key and asked it to build a genuine game — twice. Once on the model it uses by default (going direct), once through Atlas mode. Same prompt, same client, same code. Only the bill changed — and you can verify it line-by-line.

0%

lower cost across the runs — for materially the same result

$0.00
Direct
$0.00
Atlas
$0.00
Saved
8 calls·294,390 tokens·2 apps built twice

It’s one line to switch

Keep your client, your prompts, and your model choices exactly as they are. Point the base URL at Newmen, drop in your key, and Atlas handles the rest.

export ANTHROPIC_BASE_URL="https://api.newmen.ai"
export ANTHROPIC_API_KEY="nm_live_..."

Claude Code calls https://api.newmen.ai/v1/messages

Sign in to insert your real key automatically.

That’s the entire integration. No SDK swaps, no prompt changes, no model config. Every call is logged in your console with full payloads for review.

Run 1 · Heavyweight coding-agent build

A premium request, optimized by Atlas

The agent built a full Line-Rider-style physics game (draw tracks, launch a sled, gravity + collisions + particles). Going direct on the top-tier default model it billed at premium rates; with Atlas mode the identical task came in far cheaper — with a near-identical result.

The prompt

Build a single self-contained HTML file for a Line-Rider-style physics sandbox called Sled Sketch: draw tracks with the mouse, place a rider, press play, and watch a sled slide with gravity, collisions, momentum, bouncing and air. Include tool modes (draw / erase / move camera / place rider), play-pause-reset-clear, camera follow, zoom/minimap, live speed-distance-airtime-crash stats, particle effects, and best-distance saved locally — clean winter/arcade style, and genuinely playable. Then self-review the code and fix anything broken before finishing.

DirectAtlas
Cost$0.75$0.30
Calls22
Tokens89,30661,230
Throughput (tok/s)554338
Build time161s181s
Cost vs direct60% cheaper
Direct build$0.75
▶ Play it live

Polished: parallax mountains, layered glow on the track, crisp high-DPI canvas, zoom-to-fit, fine-grained collision. Best-looking of the four.

Atlas build$0.30
▶ Play it live

Feature-complete and clean: live speed bar, world grid, animated scarf, camera-follow toggle. A touch less visual depth than the direct build; plays just as well.

60% cheaper to build the same game. Trade-off seen here: the Atlas build ran ~12% slower in wall-clock — worth it for cost-sensitive or agentic workloads, and latency-critical paths can pin Strict mode per call.

Run 2 · Standard request

A standard request, optimized by Atlas

The same game build at a lighter tier. Atlas mode again cut the cost while delivering a complete, playable sandbox.

The prompt

Build a single self-contained HTML file for a Line-Rider-style physics sandbox called Sled Sketch: draw tracks with the mouse, place a rider, press play, and watch a sled slide with gravity, collisions, momentum, bouncing and air. Include tool modes (draw / erase / move camera / place rider), play-pause-reset-clear, camera follow, zoom/minimap, live speed-distance-airtime-crash stats, particle effects, and best-distance saved locally — clean winter/arcade style, and genuinely playable. Then self-review the code and fix anything broken before finishing.

DirectAtlas
Cost$0.11$0.07
Calls22
Tokens72,67571,179
Throughput (tok/s)1,292554
Build time56s128s
Cost vs direct38% cheaper
Direct build$0.11
▶ Play it live

Most configurable: track-color picker and live gravity / friction / bounce sliders. Strong UI; needs the user to draw the first track.

Atlas build$0.07
▶ Play it live

Leanest build: the core sandbox — draw, play, physics, minimap, stats. Most basic visuals of the set, but fully functional for a third of the cost.

38% cheaper for the same task. Throughput was lower than direct here — the savings come first; latency-critical paths can pin Strict mode per call.

Reading the numbers

57%
Lower cost, blended across both runs
$0.50
Saved on $0.86 of work
8
Calls, all logged for review
294,390
Tokens processed
  • Cost is the headline. 38–60% off per run, ~57% blended — with no change to the app being built.
  • Quality holds — and it's guaranteed. Every Atlas build shipped a complete, genuinely playable game: drawing tools, gravity, collisions, particles, minimap, live stats, saved best distance. The direct builds edge ahead on visual polish; the Atlas builds close most of the gap. Don't like a call? Thumbs-down it for a full refund.
  • Throughput is the trade-off. In these runs the Atlas builds were slower in wall-clock (a deliberate cost-first choice). For agentic, batch, and background work that's a clear win; latency-critical calls can pin Strict mode.
  • Effort is near-zero. Two env vars. The agent, prompts, and model names never changed.

See it in your console

None of these numbers are estimates — they’re pulled straight from your live traffic. Savings, spend, calls, tokens, and per-run breakdowns are all visible in the dashboard the moment a call lands.

Usage. Period savings, spend vs. direct, and token totals at a glance.
Overview. Verified Savings headline plus recent calls with per-call cost and savings.
Calls. Every call shows the served model, the direct price → what you paid, and the amount saved.
Sessions. Calls grouped by run, so you see what a whole task cost — and saved.

Play them yourself

Both builds in each run are embedded above — open “Play it live” on any card, draw a downhill track with the mouse, place the rider, and hit Play. The cost difference is real; judge the quality difference for yourself.

Start cutting your AI bill today.

Two env vars, your existing client, and every call logged with what it cost versus direct. Get $5 in credits and watch the savings land in your console on the first call.

Cost benchmarks · illustrative

Same task. A smaller bill — quality held.

Representative public tasks: summarization, classification, structured extraction, code review, RAG-grounded QA. Each pits the frontier baseline (going direct) against Atlas, with an evaluator deciding whether quality held. These figures are illustrative; your real, verified savings are measured on your own traffic in the console.

Total savings

$26.75

71.7% on 3,580 prompts

Baseline spend

$37.31

Frontier baselines, going direct

Atlas spend

$10.56

avg eval 0.756 vs baseline 0.786 (-3.8%)

Going direct (baseline)Atlas servedbars scaled to largest baseline

Summarization · XSum

200 prompts · rouge-1

88%cost
$11.42
$1.36
Quality heldeval 0.405 vs 0.437 (-0.032)· openai/gpt-5.5meta-llama/llama-3.1-70b-instruct via Hyperbolic

Classification · Banking77

3,080 prompts · accuracy@1

68%cost
$12.38
$3.94
Quality heldeval 0.886 vs 0.902 (-0.016)· anthropic/claude-sonnet-4.6qwen/qwen-2.5-72b-instruct via DeepInfra

Structured extraction · Invoices

50 prompts · json-schema

89%cost
$1.95
$0.21
Slight quality tradeeval 0.940 vs 0.980 (-0.040)· openai/gpt-5.5deepseek/deepseek-chat-v3 via Together

Code review · PR comments

100 prompts · llm-judge

20%cost
$5.14
$4.10
Quality heldeval 0.821 vs 0.853 (-0.032)· anthropic/claude-opus-4.7qwen/qwen-2.5-coder-32b-instruct via Hyperbolic

RAG-grounded QA · Natural Questions

150 prompts · f1-overlap

85%cost
$6.42
$0.95
Quality heldeval 0.728 vs 0.756 (-0.028)· openai/gpt-5.5meta-llama/llama-3.1-70b-instruct via Hyperbolic

Illustrative figures, quality-held-constant (eval-gated). Optimization levers such as quantization and prompt rewriting are per-knob opt-in and off by default — the mechanics live on the methodology page. Your real, verified savings are measured on your own traffic in the console, not on a benchmark.

Leaderboard

Best value. Smartest. Fastest.

Every text-generation model in the Newmen catalogue, ranked three ways and regenerated on each catalogue sync. Intelligence uses curated scores where published (everything else marked “est.”); Speed is best-of-provider p95 throughput from live traffic — the latency proof that Atlas isn't only about cost.

Latency & reliability

Not just cheaper — verified, and on the fast path when you need it.

Cost is the headline, but it isn't the whole story. Here's how Atlas keeps quality and latency honest.

Cost-first by default, fast path on demand

Atlas mode optimizes each call for the cheapest path that holds quality — a deliberate trade that can run slower in wall-clock, as the case studies show candidly. When latency rules, pin Strict mode: your exact model, sourced cheaper, on the fastest passing provider.

Verified, not asserted

Evaluators score every served output. Any call below your threshold isn't metered, and you can thumbs-down anything for a full refund. The proof lives on your own traffic — you check the math instead of taking our word.

Per-call transparency

Every call shows the served model, provider, your price, and the direct price it's measured against — reconcilable against the provider bill. Verified Savings is your own measured number versus your prior cost going direct.

Your IP stays yours

We train our optimization engine — which path holds quality and costs least — not a language model. Your prompts and code are never absorbed into a model, sold, or shared. Turn off usage-based optimization any time, or get full data isolation on Enterprise.

We only make money when you save money. If you don’t save, we don’t get paid. Period. Go deeper in the methodology or see how it works.

Talk to a solutions engineer

Atlas is sold to teams who commit to meaningful production volume. That commitment unlocks the reliability loop.