Key Terms
- AI Credit - JetBrains's billing unit for all cloud AI features including Junie: each credit equals $1 USD, quotas reset every 30 days from first use, and one credit buys roughly 10 AI Chat code-generation requests or ~40 editor code-generation requests. Token-based consumption varies by model and task length. Source: Jetbrains – Licensing And Subscriptions
- Top-up AI Credits - extra credits purchased after the monthly quota runs out; valid 12 months, available on Trial, Pro, and Ultimate (not AI Free), and shared across all licenses under one account. Source: Jetbrains – Licensing And Subscriptions
- BYOK (Bring Your Own Key) - connect your own OpenAI, Anthropic, Google, xAI, OpenRouter, or GitHub Copilot (OAuth) keys; usage is billed by the provider at its own rates with no JetBrains subscription required, and BYOK takes priority over subscription credits when a model is available both ways. Source: JetBrains – Byok
- Junie Local - on-device inference mode: run /local in Junie, the model (~15-20 GB) downloads and a local OpenAI-compatible server starts on your machine; no tokens, no quota, no code leaves the device. Source: JetBrains – Junie Local
- Junie API key (JUNIE_API_KEY) - an alternative auth method that runs Junie CLI with usage-based billing instead of a subscription; rates are not publicly listed. Source: JetBrains – Junie Cli
- Default model - Junie's recommended /model option, dynamically set by JetBrains to the model with the best price-quality ratio; which model it currently maps to is not disclosed. Source: JetBrains – Junie Cli Model Selection
- Helper models - Junie automatically uses a second model from the same provider for internal tasks (context summarization, routing, memory extraction); Anthropic primaries get a Claude Haiku helper, Google primaries get Gemini Flash, and this happens even under BYOK. Source: JetBrains – Junie Cli Model Selection
- SWE-Rebench - an independent agent benchmark run by Nebius that draws fresh tasks each cycle from recent GitHub issues to limit contamination; agents and raw models are ranked on resolved rate and pass@5. Source: Swe-Rebench
- MTP (multi-token prediction) - the speculative-decoding technique in Junie Local where a small draft model proposes tokens the main 27B model verifies in parallel; proposing 2 tokens per round made decoding 60% faster on an M5 MacBook Pro. Source: JetBrains – Smarter Local Al
- ACP (Agent Communication Protocol) - the protocol Junie CLI uses to talk to JetBrains IDEs; the same agent engine powers the IDE AI chat, the Junie tool window, and the CLI. Source: JetBrains – Junie Coding Agent Out Of Beta
Latest Changes
First report for this supplier. All models, plans, and pricing are listed as current state.
- Launch: Junie Local shipped Aug 24, 2026, running Qwen3.6-27B at 4-bit entirely on-device (about 20 GB download, requires an M5 Mac with 64 GB RAM), free with no registration or credits. Source: JetBrains – Junie Local Launch
- New model: Qwen3.8-3.6-27B-blend shipped Sep 18, 2026, a 50/50 weight merge of Qwen3.8 and Qwen3.6 that solved 37 of 100 internal coding tasks versus 34 for Qwen3.6 while generating 71% fewer output tokens than Qwen3.8. Source: JetBrains – Smarter Local Al
- Pricing promo: Gemini 3.7 Flash became the default model Aug 17, 2026 at 40% off base pricing for a limited time. Source: JetBrains – Junie Gemini 3 7 Flash
- Pricing promo: Gemini 3.8 Flash was added Sep 9, 2026 with a limited-time 75% off introductory offer in both the IDE plugin and CLI. Source: JetBrains – Junie Gemini 3 8 Flash
- Models added: Junie CLI added Grok 4.7, Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna on Sep 22, then Claude Sonnet 5.5 and GPT-6.1 Sol on Sep 29, plus a fix for false safety refusals with Claude Opus 5. Source: JetBrains – Whats New
- Feature added: /demo command shipped Sep 17, 2026, letting Junie exercise a feature in your app and produce a video, screenshots, and an HTML report attached to the diff. Source: JetBrains – Demo Agent
- Feature added: /review gained an "Apply suggestion" UI (comments posted via a new tool) and /goal got UI improvements on Sep 14, 2026. Source: JetBrains – Whats New
- Platform expansion: experimental NVIDIA RTX support landed in the Windows nightly build on Sep 18, 2026, covering Ampere-or-newer cards with at least 24 GB VRAM, on top of existing DGX Spark and RTX 5090 prototypes. Source: JetBrains – Smarter Local Al
- Milestone (context): Junie left beta on Jun 17, 2026 with Plan mode, agentic debugger control, async remote control, and context-aware code review, and was ranked the number-one coding agent on SWE-Rebench at 61.6% resolved and 72.7% pass@5. Source: JetBrains – Junie Coding Agent Out Of Beta
- Post-window: on Oct 1, 2026 (after this report's cutoff) the CLI added "Junie Lite", branded as a free AI layer with no trial and no card. Source: JetBrains – Whats New
Plans
Junie itself is free to install and run; paid plans buy cloud-model access via the JetBrains AI subscription shared across AI Assistant and Junie. Prices verified Oct 3, 2026.
| Plan | Price (monthly) | Price (annual) | Usage limits | Key inclusions |
|---|---|---|---|---|
| Free to Start | $0 | $0 | Junie Local unlimited (on-device); BYOK billed by provider | Junie CLI, IDE plugin, local models; "free to start, no credit card or subscription required" |
| AI Free | $0 | $0 | 3 AI Credits per 30 days | Cloud models in AI Chat; no top-ups; not available in Mainland China |
| AI Trial | $0 (card required) | n/a | 10 AI Credits per trial period | 30 days of AI Pro-level usage for evaluation; top-ups allowed |
| AI Pro (individuals) | $10.00 | $100.00 ($8.33/mo) | 10 AI Credits per 30 days | AI-assisted coding, AI chat, Junie; anytime top-ups |
| AI Pro (organizations) | $20.00 | $200.00 ($16.67/mo) | 10 AI Credits per 30 days per user | Same, commercial licensing |
| AI Ultimate (individuals) | $30.00 | $300.00 ($25.00/mo) | 35 AI Credits per 30 days | Higher credit quota ("recommended for Junie"); anytime top-ups |
| AI Ultimate (organizations) | $60.00 | $600.00 ($50.00/mo) | 35 AI Credits per 30 days per user | Same, commercial licensing |
| AI Enterprise | undisclosed (sales) | $60.00 per user/mo billed annually | "The maximum amount of AI credits" (undisclosed number) | Part of JetBrains IDE Services: on-premises/cloud/hybrid, SSO+SCIM, AI audit logs, zero data retention, MCP policies, BYOK |
Sources: JetBrains (free tier, credit counts, annual individual prices) , Jetbrains – Buy (full personal/commercial monthly and yearly price matrix for AIP and AIPU) , Jetbrains – Licensing And Subscriptions (AI Free/Trial credit amounts, credit mechanics) , Jetbrains – Ai Enterprise (AI Enterprise price and inclusions)
Terms explained:
- AI Credit - $1 USD of usage; roughly 10 AI Chat code-generation requests, ~40 editor code-generation requests, ~25 explain-code requests, ~280 documentation requests, or ~140 commit messages per credit, with token-based variance. Source: Jetbrains – Licensing And Subscriptions
- 30-day reset - quotas renew every 30 days from the date of first license use, not calendar months. Source: Jetbrains – Licensing And Subscriptions
- AI Ultimate bonus - 35 credits at a $30 price is a deliberate "small bonus" over the $1-per-credit correspondence. Source: Jetbrains – Licensing And Subscriptions
API Pricing
Junie is not an API supplier; it is an agent client that consumes other providers' models. There is no Junie model API to price. The billing mechanics around model usage are:
- Credit value: each AI Credit = $1 USD; subscription quotas are 10 credits/30 days (Pro) or 35 credits/30 days (Ultimate); quota consumption is "calculated dynamically for each interaction based on several factors, primarily the number of tokens used", but JetBrains publishes no per-model credit rate or token-to-credit conversion table. Source: Jetbrains – Licensing And Subscriptions
- Top-ups: extra credits can be bought anytime at $1 per credit (personal accounts via IDE widget or JetBrains Account; organizations via JetBrains Central with per-subscription limits and auto-purchase schedules); top-ups are valid 12 months. Sources: Jetbrains – Licensing And Subscriptions , Jetbrains – Licensing And Subscriptions
- Junie API key (usage-based billing): rates are not published anywhere I could verify; the docs only describe generating a JUNIE_API_KEY token. Source: JetBrains – Junie Cli
- BYOK rates: provider-rate pricing with zero markup (per the pricing card); with BYOK the request "is billed directly to the model provider without consuming credits from your JetBrains account". Sources: JetBrains , JetBrains – Byok
- Hidden helper calls: even under BYOK, Junie places calls to a second model from the same provider for routing/summarization/memory (e.g., GPT-4.1 for routing by default, Claude Haiku for Anthropic primaries, Gemini Flash for Google primaries); these appear on your provider bill. Source: JetBrains – Junie Cli Model Selection
- Promo discounts: Gemini 3.7 Flash at 40% off base pricing and Gemini 3.8 Flash at 75% off, both "limited time" with no end date or base-rate disclosure. Sources: JetBrains – Junie Gemini 3 7 Flash , JetBrains – Junie Gemini 3 8 Flash
- /usage: the CLI shows per-session cost breakdown including token usage, models used, and remaining balance. Source: JetBrains – Junie Cli
Model Performance / Benchmarks
| Benchmark | Junie (agent) | Reference entries | Notes |
|---|---|---|---|
| SWE-Rebench resolved rate (leaderboard fetched 2026-10-03, default window May 15 to Jul 1, 2026; 111 problems, 65 repos) | 61.8% ± 0.54% | Fable 5 [high] 64.5%, Grok 4.5 [high] 63.8%, Opus 5 [high] 63.4%, GLM-5.2 [high] 62.9%, GPT-5.6 Sol [medium] 62.3% | Junie ranks #6 overall; every entry above it is a raw model, not an agent harness. Source: Swe-Rebench |
| SWE-Rebench pass@5 | 73.9% | GLM-5.2 81.1%, GPT-5.6 Sol 79.3%, Claude Code 75.7% | Source: Swe-Rebench |
| SWE-Rebench cost per problem | $0.81 | GPT-5.6 Sol $0.85, Claude Code $3.39, Fable 5 $4.40 | Junie used 1,684,501 tokens per problem, 85.9% cached; cheapest per problem on the entire leaderboard among entries listed. Source: Swe-Rebench |
| SWE-Rebench (June 2026 GA cycle) | 61.6% resolved, 72.7% pass@5 | described as "top model-harness... ahead of the other agents and competitive with raw frontier models" | Quote attributed to Alexander Golubev, Research Lead at Nebius. Source: JetBrains – Junie Coding Agent Out Of Beta |
| Junie Local internal 100-task coding eval (Sep 18) | Qwen3.8-3.6-27B-blend: 37 solved | Qwen3.6 (reasoning off): 34; Qwen3.8: 39 | Blend generates 71% fewer output tokens than Qwen3.8; on the 30 tasks both completed, 279K vs 935K tokens. Source: JetBrains – Smarter Local Al |
| LiveCodeBench (avg of 4 runs, Sep 18) | Blend: 85.47% | Qwen3.8: 83.29%; Qwen3.6: 67.87% | Qwen3.6 used ~24.1M output tokens per pass versus ~6.14M for the blend. Source: JetBrains – Smarter Local Al |
| Junie Local at launch (Aug 24, private JetBrains test set) | Qwen3.6-27B "on par with Sonnet 4.5 (10,000-token reasoning limit)" | GPT-5 at medium effort scored slightly higher | Local model ships with reasoning disabled; comparison is against cloud models with reasoning on. Source: JetBrains – Junie Local Launch |
| Gemini 3.7 Flash in Junie (Aug 17) | Matched "our premium Sonnet-5 midtier model" solve rate at roughly a third of the cost per task | DeepSWE v1.1: 65.3% vs 49.0% for 3.6 Flash; FrontierCode 1.1: 43.6% vs 34.4% (Google's numbers) | Source: JetBrains – Junie Gemini 3 7 Flash |
| Gemini 3.8 Flash in Junie (Sep 9) | ~74% on DeepSWE v1.1 per Google | Outperforms most larger frontier models at significantly lower cost, per Google's chart | Number read from the post's scatter plot; spends more tokens per task than 3.7 Flash by design. Source: JetBrains – Junie Gemini 3 8 Flash |
Verification note: the homepage claim "A top performer on the SWE-Rebench benchmark" (JetBrains) checks out with the nuance that Junie is the top agent but sits below five raw frontier models on the current leaderboard.
Latest News
Gemini 3.7 Flash becomes the default at 40% off (August 17, 2026)
Google's most capable Flash model for coding replaced 3.6 Flash as Junie's default in both the IDE plugin and CLI, at 40% off base pricing for a limited time. JetBrains's private-benchmark evaluation (built from real recent commits in its own projects) found it matched its premium Sonnet-5 midtier model at roughly a third of the cost per task. Source: JetBrains – Junie Gemini 3 7 Flash
Junie Local launches, fully on-device (August 24, 2026)
Junie Local runs Qwen3.6-27B at 4-bit entirely on the machine after a single /local command and about a 20 GB download, requiring an M5 Mac with 64 GB RAM. JetBrains optimized for prefill over generation speed (the M5 Neural Accelerator's 8-bit instructions gave ~40% more prefill throughput, with a patch being upstreamed to MLX-VLM), ships speculative decoding that roughly doubles generation speed, and chose Qwen3.6 over 3.8 because 3.8 needs reasoning enabled, making tasks roughly 4x slower. Source: JetBrains – Junie Local Launch
Gemini 3.8 Flash added at 75% off (September 9, 2026)
Positioned for long multi-step jobs: it works in a loop (sub-tasks, bug-reproduction scripts, verification runs) rather than one-shot answers, spending more tokens per task in exchange for checked results. 3.7 Flash remains available and cheaper per token for small fixes. Source: JetBrains – Junie Gemini 3 8 Flash
/demo command (September 17, 2026)
Junie can now test a feature in your app and attach a video, screenshots, and an HTML report to the diff for review. Source: JetBrains – Demo Agent
Qwen3.8-3.6-27B-blend local model (September 18, 2026)
A simple equal-proportion weight merge of Qwen3.8 and Qwen3.6 (no post-training) that removes the speed-versus-capability tradeoff of the first release: more tasks solved than Qwen3.6, 71% fewer output tokens than Qwen3.8, 85.47% on LiveCodeBench across four runs, and a 24.5% theoretical token saving identified in shorter correct reasoning traces. The post also details MTP tuning (2-token proposals 60% faster; 4-token only 36%; Q4 draft head kept over Q8) and a fixed shared-random-seed evaluation bug that had artificially degraded results. Source: JetBrains – Smarter Local Al
Frontier model wave and CLI updates (September 14 to 29, 2026)
Release notes: Sep 14 adds the /review "Apply suggestion" UI and /goal improvements; Sep 22 adds Grok 4.7, Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna; Sep 29 adds Claude Sonnet 5.5 and GPT-6.1 Sol and fixes false safety refusals with Claude Opus 5. Source: JetBrains – Whats New
Context: GA and SWE-Rebench claim (June 17, 2026)
At leaving beta, Junie shipped Plan mode (structured, editable plans in .junie/plans), agentic debugging through the real IDE debugger (breakpoints in .class files and JARs), async remote control from any device, context-aware code review via GitHub Actions/GitLab (including on-prem)/CLI, and database integration through DataGrip. The SWE-Rebench number-one-agent claim (61.6% resolved, 72.7% pass@5) dates from this post. Source: JetBrains – Junie Coding Agent Out Of Beta
Community Signals
Community footprint is small relative to the product's benchmark position: HN engagement with Junie stories in the August-September window was minimal (the highest Junie Local launch thread reached 5 points), and Reddit search for Junie was blocked by a login wall this cycle (Old – Search), so no Reddit quotes are included.
Junie Local launch threads: interest with hardware and packaging pushback
The Aug 25 HN thread (5 points) drew a concrete distribution complaint. palata: "Sounds great, but I came here to complain about the installation script... Could companies stop training people to run random scripts from the internet directly in their terminal? Would it be so hard to distribute it through JetBrains toolbox and Homebrew? And how do users uninstall it then?" News – Item . A separate Aug 26 thread zeroed in on the requirement bar; zkvl quoted the spec verbatim: "you'll need an M5 Mac with 64 GB of RAM" News – Item . Third-party coverage of the launch landed on techstrong.ai (3 points, no comments): News – Item
JetBrains measurement cited in September's token-efficiency debate
In the 278-point HN thread on Spotify's claim of 90% Claude Code token savings (Sep 2026), the top-adjacent comment by Syntaf used JetBrains's own July 2026 rtk measurement as counter-evidence: "Just look at how JetBrains measured rtk and found that while an individual call saves tokens, agents on average perform *more* turns and use *more* tokens to accomplish one task" (linking JetBrains – Rtk Claude Code Token Savings ). Comment permalink: News – Item , thread: News – Item . This is a signal that JetBrains's agentic-evaluation work has credibility currency outside its own ecosystem.
Reddit recovery (added 2026-10-03)
A logged-out Reddit recovery pass (October 3) found no Junie posts above 100 votes and no Junie community in search sidebars; Reddit traction currently trails HN for this product, where the Junie Local launch thread remains the reference discussion.
Enterprise Readiness
| Feature | Available? | Details |
|---|---|---|
| SSO (SAML/OIDC) | Yes | "Advanced user access management, SSO and SCIM" on AI Enterprise. Source: Jetbrains – Ai Enterprise |
| SCIM | Yes | Listed together with SSO on AI Enterprise. Source: Jetbrains – Ai Enterprise |
| Audit logs | Yes | AI audit logs on AI Enterprise. Source: Jetbrains – Ai Enterprise |
| IP indemnity | No | Not mentioned on the AI Enterprise page, pricing pages, or any June-September post reviewed. |
| Data residency | Partial | Zero data retention on AI Enterprise; usage restricted to provider-defined allowed service territories; AI Free/Ultimate tiers not available in Mainland China. No EU-only or region-pinned inference option published. Sources: Jetbrains – Ai Enterprise , Jetbrains – Licensing And Subscriptions |
| HIPAA | No | Not mentioned on any page reviewed this cycle. |
| Air-gapped / on-prem | Yes | AI Enterprise supports "on-premises, cloud, or hybrid deployment"; Junie Local runs fully offline once weights are downloaded ("no code ever leaving your machine... works with no network at all"); CLI also connects to Ollama/LM Studio/LiteLLM. Sources: Jetbrains – Ai Enterprise , JetBrains – Junie Local Launch , JetBrains – Junie Coding Agent Out Of Beta |
| SLA | Undisclosed | No SLA terms published on pages reviewed; Enterprise is sold via contact-sales. Source: Jetbrains – Ai Enterprise |
| Admin controls (RBAC) | Yes | AI Enterprise: advanced access control, usage analytics dashboards, company-wide MCP policies; JetBrains Central organizations get a shared credit pool with per-subscription top-up limits, auto-purchase schedules, and monthly usage reports. Sources: Jetbrains – Ai Enterprise , Jetbrains – Licensing And Subscriptions |
| Compliance posture | SOC 2 | SOC 2 attestation; reports, penetration-test summaries, and security questionnaires in the Trust Center. Source: Jetbrains – Ai Enterprise |
Terms explained:
- Zero data retention - prompts and outputs are not stored by the AI service after processing; offered as an AI Enterprise control. Source: Jetbrains – Ai Enterprise
- JetBrains Central - the org-admin console where migrated organizations manage AI access grants, a shared credit pool, limits, and auto-purchase. Source: Jetbrains – Licensing And Subscriptions
Transparency Gaps
| Gap | Details | Severity |
|---|---|---|
| Per-model credit consumption | Quota is token-based and "calculated dynamically", but no table maps models to credits per task or per token; buyers cannot forecast what a Claude Opus 5.5 agent run versus a Gemini Flash run costs from a 10- or 35-credit budget. JetBrains's own docs say agent mode "leads to higher quota consumption" without numbers. Source: Jetbrains – Licensing And Subscriptions | High |
| Gemini promo terms | "40% off base pricing" and "75% off" are stated without the base rates or end dates ("limited time only"). Source: JetBrains – Junie Gemini 3 8 Flash | Medium |
| Junie API key rates | Usage-based billing via JUNIE_API_KEY is documented as an auth option, but no rate card is published. Source: JetBrains – Junie Cli | Medium |
| Default model identity | The recommended "Default" model is dynamically remapped by JetBrains and its current mapping is not disclosed. Source: JetBrains – Junie Cli Model Selection | Medium |
| Helper-model billing | Under BYOK, unselected helper models (e.g., GPT-4.1 for routing) make billable calls; the volume share of helper traffic is not quantified. Source: JetBrains – Junie Cli Model Selection | Medium |
| AI Enterprise credit amount | Marketed as "the maximum amount of AI credits" with no number; Enterprise pricing requires sales contact. Source: Jetbrains – Ai Enterprise | Medium |
| AI Free ceiling | 3 credits per 30 days with top-ups disabled locks free users out of real agent usage; Junie Local is the only unlimited free path but needs an M5 Mac/64 GB or RTX 24 GB+ nightly. Sources: Jetbrains – Licensing And Subscriptions , JetBrains – Junie Local | Low |
| Model alias drift | CLI aliases map to fixed but aging models (sonnet -> Claude Sonnet 5, opus -> Claude Opus 4.8, gpt -> GPT-5.4, grok -> Grok 4.3) that lag the models added in September; the mapping is documented but the lag is not explained. Source: JetBrains – Junie Cli Model Selection | Low |