On this page · 10 sections
Summary. Zed 1.16.1, released on 19 August 2026, lists one line under "Breaking Changes and Notices": the default model for OpenAI subscription users is now GPT-5.6 Sol. Zed bills hosted usage at "list price from the provider, +10%", and its Zed-Hosted Models page prints both numbers side by side. For GPT-5.6 Sol the arithmetic holds: OpenAI lists $5.00 input and $30.00 output per million tokens, Zed charges $5.50 and $33.00. For GPT-5.6 Terra and GPT-5.6 Luna it does not. OpenAI's live pricing page shows Terra at $2.00/$12.00 and Luna at $0.20/$1.20 per million tokens after cuts dated 30 July 2026. Zed's table still shows the older provider figures of $2.50/$15.00 and $1.00/$6.00, and bills $2.75/$16.50 and $1.10/$6.60. The published Luna rate is 5.5 times OpenAI's list price, not 1.1 times. Zed Pro costs $10 per month and includes $5 of tokens, with a default incremental spend limit of $10, so a Pro user's entire monthly AI allowance is a two-figure number that this gap moves by multiples.
What actually changed on 19 August
The 1.16.1 stable release is dated 19 August 2026. Its AI section adds Gemini 3.6 Flash to the Google models and adds Claude Opus 5, Gemini 3.5 Flash Lite, Gemini 3.6 Flash, Gemini 3.7 Flash, Kimi K3, Grok 4.6 and Muse Spark 1.2 to OpenCode Zen. Three entries sit under "Breaking Changes and Notices": the default model for OpenAI subscription users becomes GPT-5.6 Sol, OpenCode Free models are removed from Zed's built-in OpenCode provider, and the deprecated Claude Opus 4.1 model is removed from OpenCode Zen.
Read Zed's own definition before assuming this hits your invoice. The Use an Existing Subscription page separates two paths: a ChatGPT Plus or Pro subscription signs in with OpenAI and is "separate from OpenAI API keys", while Zed Pro, Business and Student seats use Zed-hosted models and are "billed through Zed". The default-model change names OpenAI subscription users, so the direct billing effect lands on ChatGPT Plus and Pro seats, whose usage is metered by OpenAI rather than by Zed.
The reason it still matters to a Zed Pro or Business team is that Sol is the expensive end of the GPT-5.6 line on every table involved, and default settings are what most engineers keep. On Zed's hosted rates, Sol output at $33.00 per million is five times Luna's $6.60. On OpenAI's own list, Sol output at $30.00 is twenty-five times Luna's $1.20. Whichever table you are billed against, moving a routine agent thread from Luna to Sol is a multiple, not a rounding error.
The price table conflict, row by row
Zed's Zed-Hosted Models page states the rule plainly and repeats it on the pricing page: "Zed Pro bills at API list price +10% for hosted usage." The Plans and Pricing doc says the same, describing usage as converted to dollars at "list price from the provider, +10%". OpenAI's pricing page, fetched 21 August 2026, gives the list.
| Model and token type | OpenAI list price per 1M | Zed's stated provider price per 1M | Zed billed price per 1M |
|---|---|---|---|
| GPT-5.6 Sol input | $5.00 | $5.00 | $5.50 |
| GPT-5.6 Sol output | $30.00 | $30.00 | $33.00 |
| GPT-5.6 Terra input | $2.00 | $2.50 | $2.75 |
| GPT-5.6 Terra output | $12.00 | $15.00 | $16.50 |
| GPT-5.6 Luna input | $0.20 | $1.00 | $1.10 |
| GPT-5.6 Luna output | $1.20 | $6.00 | $6.60 |
| GPT-5.4 cached input | $0.25 | $0.025 | $0.0275 |
Sol reconciles. Terra bills at 1.375 times OpenAI's list on both input and output. Luna bills at 5.5 times. The GPT-5.4 cached-input row runs the other way and is out by a factor of ten, with Zed's billed rate landing at about 11 percent of OpenAI's list. Cache writes drift too: Zed shows Luna cache writes at a provider price of $1.25 against OpenAI's $0.25, and Terra at $3.125 against OpenAI's $2.50.
The date explains it. OpenAI's Terra and Luna reductions took effect on 30 July 2026, alongside the rename of Priority processing to Fast mode. Zed's table carries the pre-cut numbers, which puts it 22 days behind as of 21 August 2026. Zed's Sol, GPT-5.5, GPT-5.5 pro, GPT-5.4 pro, GPT-5.2, GPT-5.3-Codex, GPT-5 mini and GPT-5 nano rows all match OpenAI exactly, so this is a stale subset rather than a different pricing policy.
What $5 of included credit actually buys
Zed Pro includes $5 of monthly token credit and defaults the incremental Monthly Spend Limit to $10, for a total of $20 per month with Zed. A trial adds $20 of credit for 14 days. The Student plan carries $10 per month and cannot configure a spend limit. Against those numbers, the model you default to decides how long the month lasts.
| Model, output tokens for $5 | At Zed's billed rate | At list price plus 10 percent |
|---|---|---|
| GPT-5.6 Sol ($33.00 billed) | 151,515 | 151,515 |
| GPT-5.6 Terra ($16.50 billed) | 303,030 | 378,787 |
| GPT-5.6 Luna ($6.60 billed) | 757,575 | 3,787,878 |
| Claude Sonnet 5 ($11.00 billed) | 454,545 | 454,545 |
| Claude Haiku 4.5 ($5.50 billed) | 909,090 | 909,090 |
A Pro user who routes agent work through Luna-class cheap models is the one carrying the gap. Sol-by-default users are billed correctly and expensively; Luna users are billed correctly against a stale table and pay roughly five times what the stated formula implies.
The Anthropic caveat that expired 11 days ago
The same Zed page carries a second stale line: "The Claude Sonnet 5 prices shown above use Anthropic's introductory pricing through August 31, 2026." Anthropic's pricing documentation now says the opposite. Its Claude Sonnet 5 row is annotated: "The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur."
Anthropic confirmed that on 10 August 2026. Zed's underlying numbers are right, at a $2.00/$10.00 provider price and $2.20/$11.00 billed, so nobody is overcharged here. The damage is planning: a team reading Zed's caveat is budgeting for a 50 percent Sonnet 5 increase on 1 September 2026 that has been cancelled. If you moved workloads off Sonnet 5 to avoid that cliff, move them back.
A third omission is harder to spot and costs real money. Anthropic states that Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text", while Claude Sonnet 4.6 and earlier use the previous one. Zed hosts both generations at rates that look flat per token. Switching from Claude Sonnet 4.6 to Claude Sonnet 5 looks like a price cut from $3.30 to $2.20 per million input tokens on Zed's table, and part of that saving is eaten by a token count that is not the same token count.
What to change this week
Set the model, do not accept it. Zed Business admins can restrict or disable hosted models by model from dashboard.zed.dev, turn off Edit Predictions org-wide, and lock data sharing settings so members cannot override them. On Pro, the Monthly Spend Limit sits on the Billing page and can be set to $0 to cap total Zed spend at exactly $10 per month. Both are the correct response to a default that changes in a weekly release.
Price your own workload against OpenAI's page, not the reseller's. That is the general lesson of this table, and it is why teams running more than a handful of seats end up putting a routing layer between the editor and the provider. Our AI gateway and model routing FinOps service exists for exactly this problem: one place where model choice, spend caps and per-team attribution are enforced, rather than in each developer's settings file. The tier-by-tier tradeoffs are covered in our guide to GPT-5.6 Sol, Terra and Luna tier selection, and the seat-versus-key arithmetic in BYOK versus subscription cost math for AI coding tools. Request-level routing, the mechanism that makes a default irrelevant, is covered in request-level model routing, and the reporting layer in the AI cost attribution ownership playbook.
Bringing your own key is the other exit. Zed supports Amazon Bedrock, Anthropic, GitHub Copilot, DeepSeek, Google AI, LM Studio, Mistral, Ollama, OpenAI, OpenRouter and Vercel keys, and spend then goes to the provider at the provider's current price. DeepSeek is a live example of why that matters: its change log records that from 16:00 UTC on 16 August 2026 the V4 family moved to peak and off-peak pricing, with off-peak set at half the peak rate. No reseller table tracks a schedule like that in real time.
India-specific considerations
Data residency is priced separately by both providers, and neither variant appears on Zed's hosted table. OpenAI charges "a 10% uplift" on regional processing endpoints for models released on or after 5 March 2026 that are eligible for data residency. Anthropic applies a 1.1x multiplier across input, output, cache writes and cache reads when inference_geo is set to us on Claude 4.6 and later, and the same multiplier on the US Data Zone Standard deployment type in Microsoft Foundry.
For an Indian team building under the Digital Personal Data Protection Act 2023, that has a practical consequence. If your data-handling position requires a pinned processing region, the price you should model is the uplifted one, and Zed's hosted rates do not publish an uplifted variant. Route those workloads through your own provider account where the residency parameter and its multiplier are both visible on the bill.
What is still unknown
Zed has not published a correction or a dated changelog entry for the pricing table, so it is not clear whether Terra and Luna are billed at the stale printed rate or at a corrected internal rate. The GPT-5.4 cached-input row is out by a factor of ten in the customer's favour, which suggests a typo rather than a policy. Zed's context table also shows GPT-5.6 Sol at 272k input and 400k total while OpenAI publishes a separate long-context tier above 272K tokens at $10.00 input and $45.00 output, and Zed's table has no long-context row for any OpenAI model. Anyone planning large-context agent runs on Zed-hosted OpenAI models should confirm the billing before committing.
FAQ
How eCorpIT can help
We build and govern AI coding-agent rollouts for engineering organisations, including the routing, spend caps and per-team cost attribution that stop a vendor's default setting from becoming a budget event. Our senior engineering teams work to CMMI Level 5 and ISO 27001:2022 practices, and we design applications aligned with DPDP requirements where processing region matters. Talk to us through /contact-us/ if your editor and agent spend is not currently attributable to a team.
References
Last updated: 21 August 2026.
