Sourcegraph

Key Terms

• The Dial - Amp's four effort modes: low (fast, small tasks), medium (default), high (deeper reasoning), ultra (hardest open-ended tasks). Current agents: GLM-5.3-Flash, Opus 5.5, GPT-6 Astra, and Fable 5.1 respectively. Source: Sourcegraph – Modes
• Orbs - Amp's cloud machines where agents run remotely and in parallel. Billed by the minute, auto-paused when idle, six sizes from a1.tiny (1 CPU, 2 GB) to a1.3xlarge (16 CPU, 44 GB), each with 60 GB disk. Source: Sourcegraph – Pricing
• Orb minutes - The unit included in paid tiers. Amp "counts minutes while an orb is running and stop[s] when it automatically pauses while idle." Megawatt includes 45,000 minutes; Gigawatt 480,000. Source: Sourcegraph – Pricing
• Runners - Amp agent processes on your own computers, a free alternative to orbs. Since September 17 one runner can serve many directories; since September 24 runners can be shared workspace-wide with --share. Source: Sourcegraph – Shared Runners
• Amp credits - Prepaid balance used for Amp-provided model inference on all models when you are not bringing your own subscription or keys. Source: Sourcegraph – Pricing
• BYOK / Model Routing - Connecting your own model subscriptions or API keys (ChatGPT subscription, OpenRouter, Amazon Bedrock, Google Cloud Agent Platform, Azure Foundry, Vercel AI Gateway, Cloudflare AI Gateway, Ollama Cloud, OpenCode Go, custom endpoints). No Amp token fees or limits on any tier except Enterprise. Source: Sourcegraph – Free Agent
• Plaid speed - A speed setting above "Fast" for GPT-6 Astra modes using OpenAI's ultrafast tier: up to 6x faster inference at 6x cost per token. Amp-provided inference only; subagents fall back to fast or standard speed. Source: Sourcegraph – Plaid Mode
• Tune Modes / Build Dial - Settings that let you pick the model and reasoning effort behind each builtin mode (main agent, Oracle, subagents) and place 2 to 4 modes, including your own plugin agents, on the Dial. Workspace admins can set a shared dial. Source: Sourcegraph – Build Your Own Dial
• Oracle - A second-opinion model the main agent can invoke. By mode it now runs GPT-6 Astra (low, medium, ultra) or Fable 5.1 (high) at x-high effort. Source: Sourcegraph – Modes
• Subagents - Specialized agents with their own context: Search (GPT-5.6 Terra), Librarian (GPT-5.6 Sol), Read Thread (GLM-5.2). Source: Sourcegraph – Modes
• Puck - Amp's meta-agent helper for navigating projects and threads, powered by GPT-5.6 Sol, with realtime voice via GPT Live 1. Source: Sourcegraph – Modes
• Zero and minimal data retention (ZDR) - Provider-side policy that Amp usage is not stored or trained on. Since September 13 it is contractually guaranteed to all users and workspaces, not just Enterprise. Source: Sourcegraph – Free Agent

Latest Changes

Changes since the 2026-08 report.

• Pricing overhaul: "Free Agent" (September 13) made Amp free with your own compute and model subscriptions/keys: a new $0 Hobby tier, Teams at no extra charge, no monthly plan required to use a ChatGPT subscription, and no BYOK token fees or limits for anyone except Enterprise. Source: Sourcegraph – Free Agent
• Subscription contents: Megawatt stays $20/month but is now sold as 45,000 orb-minutes (750 hours) and Gigawatt stays $200/month with 480,000 orb-minutes (8,000 hours); Amp says members average 60% (Megawatt) and 65% (Gigawatt) discounts on Amp usage. The dollar usage allowances shown in August are no longer published. Source: Sourcegraph – Pricing , Sourcegraph – Free Agent
• Model change: Claude Opus 5.5 replaced GPT-5.6 Sol as the medium agent (September 28): 65% internal solve rate versus 61% for GPT-5.6 Sol, at 10% lower cost; it runs at high effort because xhigh/max cost several times more and score lower. Source: Sourcegraph – Opus 5.5
• Model change: GPT-6 Astra now powers high (the default High mode) at medium effort, with Oracle on Fable 5.1; Amp evaluated GPT-6 Sol and rejected it for medium as "jagged" despite matching GPT-5.6 Sol's score at half the cost. Source: Sourcegraph – Modes , Sourcegraph – Opus 5.5
• Model change: low switched from GLM-5.2 to GLM-5.3-Flash (high effort), and the Read Thread subagent remains on GLM-5.2. Source: Sourcegraph – Modes
• Cost change: Fable 5.1 has powered ultra since September 1: ultra threads are about 35% cheaper than on Fable 5, cache reads are 75% cheaper, and over 90% of tokens in a typical Amp thread are cache reads. Source: Sourcegraph – Fable 5.1
• New capability: Plaid speed (September 29) gives GPT-6 Astra modes up to 6x faster inference at 6x cost per token, on Amp-provided inference only (OpenAI does not yet support ultrafast for linked ChatGPT subscriptions). Source: Sourcegraph – Plaid Mode
• New capability: Customize Your Dial (September 10) adds Tune Modes (per-mode model/effort overrides for main agent, Oracle, and subagents, billed to your own keys or subscription) and Build Dial (2 to 4 modes, including plugin agents, with a workspace-shared dial option). Source: Sourcegraph – Build Your Own Dial
• Runners maturation: Four releases in eight days: one runner serving many directories (September 17), Git worktrees/projects from the thread composer plus workspace secrets (September 22), the Mac app as an always-on runner (September 24), and workspace-shared runners via --share (September 24). Source: Sourcegraph – Shared Runners
• Enterprise/security: Zero and minimal data retention was contractually extended to all users (September 13); the Enterprise card now lists audit logs, network/MCP access controls, spend limits and reporting, and IP allowlisting; SAML/OIDC SSO is free for any workspace with at least one paid member. Source: Sourcegraph – Free Agent , Sourcegraph – Pricing

Plans

Amp restructured its plans on September 13 around "bring your own tokens, pay Amp for compute." Subscriptions buy orb minutes and Amp credits; models can come from your own subscriptions or keys on every tier. Source: Sourcegraph – Pricing , Sourcegraph – Free Agent

Terms explained:

• SAML/OIDC SSO - Employees log in through their corporate identity provider; Amp offers it free for the whole workspace if at least one member is on a paid tier. Sourcegraph – Free Agent
• SCIM - Automatically provisions and deprovisions Amp accounts when employees join or leave. Sourcegraph – Pricing
• Pooled credits - A workspace-wide balance that admins fund and members consume for orbs pay-as-you-go, so most of the team can stay on the free Hobby tier. Sourcegraph – Free Agent
• Forward-Deployed Engineering - Amp Labs engineers embedded with enterprise customers; listed as available on the Enterprise plan. Amplabs

Notes:

• A paid tier "includes orbin' time (minutes of orbs, our machines in the cloud that run your agents) and Amp credits (which you can use for inference on all models)." Source: Sourcegraph – Pricing
• Orb minutes are counted "while an orb is running" and stop when the orb auto-pauses while idle. Source: Sourcegraph – Pricing
• All users and workspaces, not just Enterprise, are covered by the zero and minimal data retention policy; the pricing Q&A also references SOC 2 Type 2 certification. Source: Sourcegraph – Pricing
• Education discount: students and teachers get Megawatt for $10/month. Source: Sourcegraph – Pricing

API Pricing

Amp does not sell a standalone model API. For Amp-provided inference it states "no markup" but publishes no per-model $/MTok rates; the only prepaid instrument is Amp credits. When you use your own keys or subscriptions (via Model Routing or Tune Modes), Amp charges nothing and imposes no token limits, except on the Enterprise tier, and "billing follows your provider connections." Source: Sourcegraph – Free Agent , Sourcegraph – Build Your Own Dial

Current model roles (rates not published by Amp):

Source: Sourcegraph – Modes

Orb pay-as-you-go rates (published on the pricing page):

Source: Sourcegraph – Pricing

Cost data points from Amp's own posts (September 2026):

• Opus 5.5 (medium): solves tasks for 10% less than GPT-5.6 Sol and 25% less than Opus 5. Source: Sourcegraph – Opus 5.5
• GPT-6 Sol: matches GPT-5.6 Sol's score at half the cost, but Amp found it inconsistent on real work. Source: Sourcegraph – Opus 5.5
• Fable 5.1 (ultra): threads about 35% cheaper than Fable 5; cache reads 75% cheaper; over 90% of tokens in a typical thread are cache reads. Source: Sourcegraph – Fable 5.1
• Plaid: 6x cost per token versus standard, for up to 6x faster inference on GPT-6 Astra. Source: Sourcegraph – Plaid Mode
• Subscription value: average 60% discount on Amp usage for Megawatt members, 65% for Gigawatt. Source: Sourcegraph – Free Agent

Important: The absolute dollar cost of a thread on Amp-provided inference still cannot be computed ahead of time from Amp's own pages, because no per-token base rates and no Amp-credit amounts per tier are published. Per-thread costs remain visible after the fact in the product (the August "Explain Usage" feature lets you ask Puck where tokens went, listed on the Chronicle). Source: Sourcegraph – News

Model Performance / Benchmarks

Amp publishes no public benchmark scores (SWE-Bench Verified, Terminal-Bench, etc.). It reports internal eval solve rates in launch posts; the task set, prompts, and pass criteria are not published. Source: Sourcegraph – Opus 5.5

The eval methodology is not published, so Amp's internal percentages are not comparable to vendor SWE-Bench or Terminal-Bench submissions.

Latest News

Fable 5.1 Powers Ultra (September 1, 2026)

Claude Fable 5.1 replaced Fable 5 in ultra. Amp highlights long-horizon proof of work (it iterates until its evidence holds), better mid-task steering, and writing quality; ultra threads are about 35% cheaper, with cache reads 75% cheaper and over 90% of typical-thread tokens being cache reads. Cited internal results: Safari typing latency in the iOS app cut from 85 ms to 8 ms, and ampcode.com thread creation made 45% faster. The same day, Amp shipped Intelligently Ordered Diffs, which puts files that explain a change first in diffs. Source: Sourcegraph – Fable 5.1 , Sourcegraph – News

Desktop in Orbs (September 4, 2026)

Any orb thread now has a Desktop tab providing a high-resolution interactive Linux desktop, for verifying agent work targeting Linux, Android, or the Chrome extension API, opening file formats in desktop apps, or general computer use. Source: Sourcegraph – Desktop

Customize Your Dial (September 10, 2026)

Two new settings tabs: Tune Modes lets you set the model and reasoning effort for each builtin mode's main agent, Oracle, or subagents, run through your own API key or ChatGPT subscription ("billing follows your provider connections"); Build Dial lets you drag 2 to 4 modes, including plugin agents, onto the Dial, with an optional workspace-wide shared dial. Shipped two days after "Steer, Don't Queue" (September 8) made Enter send steering messages to the agent mid-task instead of queueing them. Source: Sourcegraph – Build Your Own Dial , Sourcegraph – News

Free Agent: Amp's Pricing Rebuild (September 13, 2026)

Amp became free to use with your own compute and model subscriptions/keys: no monthly plan is required to use a ChatGPT subscription with Amp, and there are no BYOK token fees or limits except on Enterprise. Amp now sells orbs (remote compute) and, optionally, no-markup inference via Amp credits. New free Hobby tier; Teams at no extra charge with free SAML/OIDC SSO when at least one member pays; Megawatt and Gigawatt members keep "heavily discounted orb usage and model usage" (60%/65% average discounts) and are no longer bound by ChatGPT subscription limits. Zero and minimal data retention was contractually extended from Enterprise to all users. A September 30 update opened early-access BYOK providers to every tier: OpenRouter, Amazon Bedrock, Google Cloud Agent Platform (formerly Vertex AI), Azure Foundry, Vercel AI Gateway, Cloudflare AI Gateway, Ollama Cloud, OpenCode Go, and custom endpoint URLs. Source: Sourcegraph – Free Agent

Runner Month (September 17 to 24, 2026)

Four releases made local machines first-class execution targets: "One Runner Is Now Enough" (September 17) lets a single runner serve many directories and discover Git repositories; "One Runner, Many Worktrees" (September 22) adds worktree/directory/project creation from the thread composer plus workspace and project secrets; "The Mac App Is Your Runner" (September 24) turns the macOS app into an always-on runner; "Shared Runners" (September 24) lets any member start threads on a --shared machine from ampcode.com. Amp warns that shared-runner users "run code on your machine as you, with your files, your credentials, and your logins," and admins can disable sharing workspace-wide. Source: Sourcegraph – Shared Runners

Opus 5.5 Powers Medium (September 28, 2026)

Claude Opus 5.5 replaced GPT-5.6 Sol as the default medium agent after solving 65% of internal eval tasks versus 61%, at 10% lower cost. ChatGPT Only preset users keep GPT-5.6 Sol billed to their subscription. Amp documents steering behavior (it folds mid-task corrections into ongoing work), and states Opus 5.5 runs at high effort because beyond high "it burns through tokens, scores lower, and costs several times as much." Amp also disclosed it ran GPT-6 Sol (shipped by OpenAI an hour after Opus 5.5) and found it equal-scored at half cost but too inconsistent for the default mode. Source: Sourcegraph – Opus 5.5

Plaid Speed, and BYOK for Everyone (September 29 to 30, 2026)

Plaid mode puts GPT-6 Astra modes on OpenAI's ultrafast tier: up to 6x faster inference at 6x cost per token, toggleable per thread, Amp-provided inference only. Subagents and non-Plaid models fall back to fast or standard speed. On September 30, Amp moved its expanded BYOK provider list from paid-tier early access to all users on every tier, configured in Model Routing. Source: Sourcegraph – Plaid Mode , Sourcegraph – Free Agent

Community Signals

The "Orbs" essay was Amp's biggest HN moment of the year (116 points, 43 comments)

Thorsten Ball's August 25 essay "Orbs, Explained" ( Amp's own reflections on why remote agent machines changed its workflow ) reached 116 points and 43 comments on HackerNews, with discussion running from August 30 into September. It remains Amp's highest-engagement HN story since the December 2025 spinout. The thread split on whether orbs are a real product category or rebranded VMs. Source: News – Item

• fishtoaster dismissed the category: "Same thing as: 'Cloud Agents' on cursor, 'Cloud Coding Agents' on Hoplite.sh, 'Sessions' in Devin cloud (they don't make a big deal of them being remote). Amp is just trying to pitch an increasingly-commodity experience as a special sauce." News – Item
• misternugget (Ball, the post's author) replied: "Yes, like many things, Orbs are 'wrappers' around VMs, but the point is that some wrappers enable new ways to hold and use and reuse the thing they wrap... some people say the tortilla is what makes it a burrito, you know?" News – Item
• altschuler countered with a paying user's view: "I've found Orbs to be just fantastic!... It's the best $20/month I'm spending on any dev tools, the Amp coding agent in itself is worth that much (I actually came for that and stayed for the orbs)." News – Item
• josevalerio compared providers: "OpenAI already does this with Codex cloud... From my experience, Orbs have been more consistent, along with Cursor's agents." News – Item
• aarondong warned on consolidation: "It sounds like great engineering, but everyone in the coding agent space is building the same thing. And big labs like Anthropic and OpenAI have more money to burn." News – Item
• sqs (Amp team) responded to a signup-flow complaint in-thread: "I am on the Amp team. I just turned off SMS verification. Not needed anymore. Was required for ~5% of signups for anti-abuse, but I think we're good without it now." News – Item

The Free Agent pricing change drew little HN attention, and skepticism about the pivot (12 points, 2 comments)

September's biggest business news, the September 13 free-tier restructure, was submitted on September 12 and reached only 12 points with 2 comments. Source: News – Item

• pjm331: "I've always been surprised about the lack of discussion of amp on the front page here. Tough selling dev tools these days. It feels like the Venn diagram of people who would use orbs and people who will just implement orbs themselves on top of something like E2B is basically a circle." News – Item
• gapo read the move as a strategy shift: "Surprised they're pivoting from the harness to the agent compute? Now. Always thought they were a great harness fwiw." News – Item

General observations

• A date-filtered HN search for sourcegraph stories since June 14 returned zero results; Amp's community footprint now runs entirely under the Amp brand. Source: Hn – Search By Date
• No September HN story covered the Opus 5.5 switch, Plaid, or the Dial customization; as in June, Amp's model-adoption speed is not mirrored by community discussion volume.
• Reddit search for sourcegraph+amp was blocked by a login wall and could not be used this cycle. No X/Twitter community data was collected.

Reddit recovery (added 2026-10-03)

A logged-out Reddit recovery pass (October 3) surfaced no Amp posts in the month's top results and no Sourcegraph/Amp community in search sidebars; Amp's community footprint remains HN-native (the "Orbs" and "Free Agent" threads).

Enterprise Readiness

Terms explained:

• Network/MCP access controls - Enterprise rules governing which MCP servers and network destinations agents may reach. Sourcegraph – Pricing
• Zero and minimal data retention - Amp's provider-side commitment that usage is not stored or used for training, now contractually guaranteed to all tiers. Sourcegraph – Free Agent

Transparency Gaps