NEO
Autonomous AI engineering agent that runs ML experiments, LLM evals, fine-tuning and prompt optimization from a single task prompt.
Only worth your time if you are closing out an existing project before the October 30, 2026 shutdown and still hold credits. Up to that point NEO was one of the few agents we've seen actually finish the job — 34.2% and first place on MLE-bench in August 2025 — and its blog published negative results, not just wins. For anything new, start elsewhere.
Verified 2d ago · liveness 70/100 · cite: rightaichoice.com/tools/neo
- ML engineers automating fine-tuning, evaluation and experiment sweeps
- Researchers benchmarking LLMs across large task suites with reproducible artifacts
- Teams finishing RAG pipeline and prompt optimization work before the shutdown
- Practitioners running agents and agent swarms on their own GPU compute under a VPC
- Complete beginners without an ML or Python background
- Anyone needing a drag-and-drop, no-code ML interface
- One-off scripts where agent setup costs more than writing the code yourself
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip NEO if you are starting anything new rather than finishing an existing run — the platform shuts down permanently on October 30, 2026 at 11:59 PM PDT and billing has already stopped, so there is no ongoing roadmap to build on.
Cheap tiers are gated by mode: Starter ($29/mo) and Value ($69/mo) are Lite mode only, so anyone needing Pro mode work has to jump to Pro at $199/mo.
Priced for working ML teams, not hobbyists: $29/mo Starter (29K credits, Lite only), $69/mo Value (100K credits, Lite only), $199/mo Pro (250K credits plus 720 hours of dedicated compute) and custom Enterprise with SSO and audit logs. The real comparison is against RD-Agent and AIDE, both of which are cheaper on licence cost but expect you to build the orchestration yourself. Subscription billing has stopped ahead of the October 30, 2026 shutdown.
In short
NEO — Autonomous AI engineering agent that runs ML experiments, LLM evals, fine-tuning and prompt optimization from a single task prompt. Best for ML engineers automating fine-tuning, evaluation and experiment sweeps, Researchers benchmarking LLMs across large task suites with reproducible artifacts, Teams finishing RAG pipeline and prompt optimization work before the shutdown. Free to start; paid plans from $29/mo.
What's new in NEO
Checked todayAcross the latest 9 updates: 2 changelog entries and 7 news mentions.
Neo to shut down permanently on October 30, 2026
Neo will permanently shut down October 30, 2026 at 11:59 PM PDT. Subscription billing has stopped; existing credits remain usable until the deadline.
Neo research loop lifts pi coding agent from 82% to 95% on QuixBugs
Neo ran an autonomous research loop on the pi coding agent, raising QuixBugs from 49/60 to 57/60 on development and 19/20 on a sealed holdout with one file changed.
Neo benchmarks Qwen3.8, Qwen3.6, and Muse Glimmer on grading divergence
On 38 comparable answers, Muse Glimmer 30B won deterministic checks while a blind LLM judge preferred Qwen3.8-27B. The winner depended on grading method.
Neo benchmarks CPU voice cloning: Pocket TTS, Kokoro, Audio8, XTTS-v2
Audio8 matched XTTS on identity (SIM 0.5454 vs 0.5569) and led among cloners on UTMOS 3.3657; Kokoro kept the lead in fixed-voice quality and speed.
Neo study compares Grok 4.6, Kimi K3, Opus 5 on Kokoro-82M speedups
Opus scored 83, Kimi 60, Grok 51. Grok's best verified gain was about 6-8% in one setup, but audio checks failed, so nothing shipped.
Neo makes OmniVoice about 6x faster than its 32-step default
Neo reports roughly 6x speedup over OmniVoice's default path and about 3x over the official fast setting, without retraining and with quality checks retained.
Neo ASR benchmark: five models tie on clean speech, diverge on accents and noise
Across 296 real clips and 19 conditions, clean-speech WER tied at 4.4%, with accent and noise separating the field. Neo committed transcripts, bootstrap CIs, and a CI regression gate.
Neo compares Kimi K3 and Opus 5 on Kokoro-82M CPU optimization
Using Neo BYOK, Opus built the stronger study but its claimed 18.66% speedup became 9.81% slower on replay.
Neo benchmarks Kimi K3, GLM 5.2, Fable 5 on ML engineering artifacts
Three AutoML artifacts passed all suites (45/45, 83/83, 37/37), but execution found silent trust failures. Scores: Kimi 68, Fable 61, GLM 54.
What people actually say about NEO — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
44 mentions across 4 sources (Hacker News, GitHub, Lemmy, Tech Press) · researched Jul 2, 2026.
Average across the 4 sources that answered — each source counts once, not each post.
- +Designed to automate end-to-end ML workflows from a single prompt.
- +Supports self-correction and multi-step reasoning loops for iteration.
- +Integrates with VS Code, Cursor, and popular LLMs.
- +Offers a free trial to test capabilities before committing.
- +Enables prompt optimization and LLM fine-tuning via natural language.
- −No real user feedback available to validate advertised features.
- −Credit-based pricing can lead to unexpected costs if experiments run long.
- −Limited platform support — only VS Code and Cursor plugins mentioned.
- −No independent reviews or case studies to back up marketing claims.
- −Name confusion with other products makes research difficult.
- • Credit consumption may be higher than expected for complex workflows.
Viability Score
How well maintained and how widely used is NEO? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Autonomous ML engineering from a single natural-language task prompt
- Multi-day autonomous runs returning versioned artifacts for review
- LLM evaluation and benchmarking across 150+ real-world tasks in 10 categories
- Dual-LLM prompt optimization loop with a JSON ledger tracking every iteration
- RAG pipeline building, evaluation and optimization
- Fine-tuning and training of ML models and LLMs
- Agent-swarm framework coordinating specialized agents over an async message bus
- GPU sandbox running hundreds of experiments and auto-selecting the best model
- Interactive chat to steer NEO's exploration of models and approaches
- VS Code extension with inline chat, plan view and file explorer
- Cursor extension support
- Open VSX extension support
- MCP server for driving NEO from Claude Code or Codex
- Bring your own LLM (BYOK) to power the agent
- Bring your own compute (VPC) with CPU/GPU and no quota limits
About NEO
NEO (heyneo.com) is an autonomous AI engineering agent that takes a plain-language task — fine-tune a model, build a RAG eval pipeline, benchmark an LLM, optimize a prompt — then researches approaches, writes the code, runs experiments, debugs failures and returns versioned artifacts you can review and promote. Runs can stretch for days, and you steer through chat whenever the plan drifts. It is built for ML engineers and researchers who already know Python and have opinions about compute, not for people looking for a no-code dashboard. There are three ways to run it: a VS Code, Cursor and Open VSX extension with inline chat, plan view and file explorer; an MCP server so you can drive NEO from Claude Code or Codex; or a cloud mode with no setup. You can point it at your own CPU/GPU compute under a VPC, which removes quota limits and keeps spend on your infrastructure. Lite and Pro modes split lighter IDE work from heavier autonomous runs. The machinery underneath is specific: a dual-LLM optimization loop where an optimizer LLM writes prompts and reads failure summaries while a target LLM runs batches against synthetic data, with a JSON ledger tracking each iteration until scores converge; an agent-swarm framework coordinating specialized agents over an async message bus; and a GPU sandbox that runs hundreds of experiments and auto-selects the best model. Benchmarking covers models from OpenAI, Anthropic, Google and more across 150+ real-world tasks in 10 categories, and NEO's public record includes a 34.2% score and first place on MLE-bench in August 2025, ahead of RD-Agent and AIDE, with 75 competitions entered. One thing overrides everything else: NEO's site carries a shutdown notice — the platform permanently shuts down on October 30, 2026 at 11:59 PM PDT, subscription billing has stopped, and existing credits can be used until then. Against alternatives like open lab frameworks or hosted eval platforms, NEO was the autonomous-loop option — but it is now a
Behind the Verdict
Buy this only if you have unfinished work in flight. NEO's own site says the platform permanently shuts down on October 30, 2026 at 11:59 PM PDT, subscription billing has stopped, and you can burn existing credits until then. That single fact reframes every feature review: the question is no longer whether NEO is good, it is whether your remaining runs finish in time. Within that window, the autonomous loop is real. You state an outcome, point NEO at your repo, data and constraints, and it writes code, runs long experiments, evaluates and hands back versioned artifacts. Runs can last days, and you intervene through chat when the plan drifts. Where it bites: credits are the throttle. Starter's 29K credits a month on Lite mode only will not carry a long fine-tuning sweep, so anyone doing serious work is looking at Pro's 250K credits plus 720 hours of dedicated compute, or Enterprise for custom allocation and Pro model access. The compute model is the part we'd flag for buyers. Every paid tier includes bring-your-own-compute under a VPC, which removes quota limits but makes you responsible for provisioning CPU or GPU capacity. Pro adds platform-hosted dedicated compute; Enterprise is the only tier with VPC plus dedicated platform compute, SSO, audit logs and an SLA. If you have no GPU infrastructure and no appetite to build it, this is not your tool. The closest non-NEO route is running an open agent framework on your own cluster, or a hosted eval and experiment platform with a UI. Those give you a future. NEO gives you a proven autonomous loop with documented results — the pi coding agent lifted from 49/60 to 57/60 on QuixBugs dev and 19/20 on a sealed holdout with the same model and five prompt guidelines, an OmniVoice speedup of roughly 6.2x over its 32-step default
Researching NEO? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas NEO actually fits — and what changes day-one when you adopt it.
You point NEO at your repo and dataset, describe the fine-tuning goal, and let it research approaches and run sweeps. It writes the code, runs hundreds of experiments in the GPU sandbox, evaluates each result and returns versioned artifacts.
Outcome: You review the winning run's artifacts, replay it on a holdout, and promote it — with credits as the only metered cost because compute runs on your own VPC.
You use NEO BYOK with two different driving models and ask it to build the same evaluation harness twice, then compare the artifacts on replay rather than on their test suites.
Outcome: The execution-inverted rankings in NEO's own Kimi K3 / GLM 5.2 / Fable 5 study (68 / 54 / 61) replicate in your domain, catching artifacts that passed their tests but failed in practice.
You hand NEO a failing eval set and let it iterate with the dual-LLM prompt loop, reading failure summaries from the JSON ledger each round until scores converge.
Outcome: You get optimized prompts plus a written record of every iteration, and can decide whether the improvement is worth promoting to staging.
Use Cases
- Automate fine-tuning of a large language model on a custom dataset with iterative hyperparameter optimization.
- Build and evaluate a multi-agent trading system with async coordination over real market data.
- Benchmark multiple LLMs across coding, reasoning and structured output tasks using a unified framework.
- Synthesize and govern a high-quality dataset for agent failure analysis with provenance and validation.
- Optimize prompts for a production chatbot using the dual-LLM loop and synthetic data batches.
- Deploy an end-to-end speech-to-speech pipeline by delegating research and implementation to NEO over MCP.
- Benchmark voice cloning and TTS models on CPU to identify optimal quantization for target hardware.
- Compare frontier models on production ML engineering tasks where green test suites can hide silent failures.
Models Under the Hood
as of 2026-10-15
Limitations
- NEO runs on credits: Starter gives 29K credits/month, Value 100K, Pro 250K, Enterprise custom.
- Starter and Value are Lite mode only, and Pro Models access requires Pro or Enterprise.
- Starter, Value and Pro require bringing your own compute via VPC — dedicated platform compute is Enterprise-only — so the list price understates real cost unless you already have CPU/GPU capacity.
- Because runs are autonomous and can last for days, exhausting credits can interrupt long experiments, and the tool is practitioner-oriented, so expect a learning curve around credit budgeting, compute configuration and steering the agent.
- The overriding constraint is the shutdown: NEO's own site states the platform permanently closes on October 30, 2026 at 11:59 PM PDT, subscriptions have stopped billing, and remaining credits are only usable until then.
- Any workflow you build on NEO needs a successor path before that date.
as of 2026-10-09
Verification history
We have re-verified NEO 8 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 8 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published NEO tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Starter
$29/mo
Ideal for
A solo ML engineer testing autonomous runs on their own VPC compute without a team budget
What this tier adds
Starting tier: 29K credits/month, Lite mode only, community support, and you must bring your own CPU/GPU compute via VPC
Value
$69/mo
Ideal for
An individual practitioner or small team running regular eval and prompt-optimization jobs
What this tier adds
Adds 71K more monthly credits over Starter (100K total) and basic email support, still Lite mode only and still BYO compute
Pro
$199/mo
Ideal for
A working ML team that needs Pro mode for heavier autonomous runs and doesn't want to hit credit ceilings mid-experiment
What this tier adds
Adds Pro mode alongside Lite, 250K credits/month, 720 hours of dedicated compute, Pro Models access, concierge onboarding and priority email support
Enterprise
Custom
Ideal for
Larger organisations needing SSO, audit logs, an SLA and dedicated platform compute rather than bringing their own bare metal
What this tier adds
Adds custom credit allocation, dedicated platform compute, a dedicated success manager, SSO and audit logs, and an SLA guarantee
Where the pricing makes sense
The company stage and team size where NEO's pricing actually pencils out — and where peers do it cheaper.
Priced for working ML teams, not hobbyists: $29/mo Starter (29K credits, Lite only), $69/mo Value (100K credits, Lite only), $199/mo Pro (250K credits plus 720 hours of dedicated compute) and custom Enterprise with SSO and audit logs. The real comparison is against RD-Agent and AIDE, both of which are cheaper on licence cost but expect you to build the orchestration yourself. Subscription billing has stopped ahead of the October 30, 2026 shutdown.
Setup time & first value
How long it actually takes to get something useful out of NEO — broken out by persona, not the marketing-page minute.
Installing the VS Code, Cursor or Open VSX extension takes a few minutes, and cloud mode avoids local setup entirely. First useful value lands inside a single short task. Budget longer — most of a day — if you route runs through your own VPC, because that's where compute and credential configuration actually lives. Teams working past the October 30, 2026 shutdown should budget that time for
Integrations
Resources & Guides
Tutorials & Learning
YouTube returned 6 videos for “NEO”, and we withheld 6: 6 could not be judged, because “NEO” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about NEO.
Official links
Tools that pair well with NEO
Common stack mates teams adopt alongside NEO, with the specific reason each pairing earns its keep.
Open Interpreter
Open Interpreter is an open-source terminal coding agent that turns plain-English requests into real file edits and shell commands.
Codex by OpenAI
OpenAI Codex: a cloud coding agent that runs many software engineering tasks in parallel
OpenHands
OpenHands runs autonomous coding agents triggered by GitHub, Slack, Jira, or schedules to review PRs and fix CI.
Featured Head-to-Head Comparisons
Neo vs Locus Robotics
Locus Robotics and NEO serve completely different domains – warehouse logistics vs. ML automation. Pick Locus if you need to physically move goods faster in a high-volume fulfillment center; its new Locus Array (2026) pushes autonomous picking with Physical AI. Pick NEO if you build ML models and need an autonomous agent to run experiments, fine-tune LLMs, and build RAG pipelines – its June 2026 BYOK capability lets you use your own LLM for identical workflows. There is no overlap; choose based on whether your problem lives in the physical or software world.
Neo vs Presto Voice
Presto Voice and NEO serve completely different domains: Presto automates drive-thru ordering for QSR chains, while NEO is an autonomous agent for ML engineering. Choose Presto if you run a multi-location drive-thru restaurant looking to boost revenue through AI upselling; choose NEO if you're an ML engineer needing an autonomous system to build, evaluate, and optimize models and agents.
Neo vs Truleo
Truleo and NEO are purpose-built for entirely different domains—law enforcement intelligence vs. ML engineering automation. Your choice hinges on your sector: pick Truleo if you're a police agency needing to unify siloed data and automate lead generation, or NEO if you're a machine learning engineer seeking an autonomous agent for model development, evaluation, and optimization. Both are powerful in their niches, but cross-over is nonexistent.
Alternatives to NEO
View all
Open Interpreter
Open Interpreter is an open-source terminal coding agent that turns plain-English requests into real file edits and shell commands.
Codex by OpenAI
OpenAI Codex: a cloud coding agent that runs many software engineering tasks in parallel
Frequently Asked Questions
Best-of guides
Used NEO? Help shape our editorial sentiment research.