Beam 1.0 Is Free Inside OpenCode: 1M Context, Unlimited Requests, and a Free Window With No Expiry Date

Reflection's Beam 1.0-501B-A23B became free to call inside OpenCode on 16 October 2026 — 1,048,576 tokens of context, no request ceiling printed anywhere, and a "limited time" label with no end date attached to it. The model itself is from early October — the catalogue entry behind OpenCode's picker carries a 5 October 2026 release date — so what changed this week is where you can reach it and what it costs to reach it, which is a narrower piece of news than "free model appears in a coding harness"; that happens most weeks on this platform.

Two details decide whether that matters to you or is merely interesting. The first is retention. Beam Free is not one of the free models whose provider trains on your prompts, but it is not zero-retention either: OpenCode publishes 30 days. In a coding agent the payload is your repository, which turns that line into a decision about your source code rather than about model quality. The second is price. Beam 1.0 is not free everywhere: the Apache 2.0 weights are served commercially by third-party providers, and Artificial Analysis records a median of $0.70 per million input tokens and $2.10 per million output tokens across the providers it tracks. That is an independent reading rather than Reflection's, and it is the only independent price reading this model has: third-party endpoints have rate cards of their own, Reflection has none, so what Reflection itself will charge once the free window closes is not something anyone can read today.

What OpenCode actually published

The announcement is a single line. OpenCode's own post on X, dated 16 October 2026, reads: "Beam is now available in OpenCode - 1M Context - Free for a limited time". Everything checkable about the offer sits behind that sentence, in two places that publish different things.

OpenCode's Zen documentation is the authority on the endpoint. It lists the model as beam-free, served at https://opencode.ai/inference/openai/v1/chat/completions through the O​penAI-compatible @ai-sdk/openai-compatible package — meaning an existing O​penAI-shaped client points at it with a base-URL change and nothing else. The pricing row is Free on input, Free on output, Free on cached read, and a dash in the cached-write column — the same dash every free row on that table carries. That dash is worth noticing: on this route the model is free to read from and free to write to, so there is no partial meter hiding in the offer.

The Go page is the authority on the limits, and it prints no numbers. Beam Free appears there tagged "limited time" with "Unlimited" in the monthly-usage column — identically on the $10 Go plan and the $40 Go Plus plan — and "Unlimited" against all three request-estimate columns in the estimated-requests table. LongCat 2.5 Preview Free carries the same treatment; nothing else in the list does. The Go model list also flags Beam Free as "limited regions", which is a limit of a different kind and is dealt with below. OpenCode's page states that its models are hosted in the US and EU.

The model metadata behind OpenCode's picker adds the specs. The entry under reflection/beam-1.0-501b-a23b carries a 1,048,576-token context window, a 131,072-token maximum output, and yes-flags on reasoning, tool calling, structured output and temperature control. Weights are marked open, pointing at Reflection's Hugging Face repository. The price field reads $0.00 / $0.00 — which is not Reflection's rate. It is the absence of one, recorded the only way a catalogue can record it while a free route exists.

The retention line is doing more work than the price line

OpenCode publishes per-model privacy terms on the same page, and in a table where "free" appears fifteen times, the retention column is the part that actually separates the offers. Read the free models against each other and they sort into three groups.

• Trains on your data during the free period — Big Pickle, Exo Free, MiMo-V2.6-Flash Free and MiMo-V2.5 Free, where OpenCode states plainly that collected data "may be used to improve the model". M​use Spark 1.3 Contributor Free goes further in a different direction: the discount is granted "in exchange for permission to use your prompts and completions to train future M​eta models".

• Zero retention, no training — LongCat 2.5 Preview Free, GLM-5.3-Flash Free, DeepSeek V4.1 Flash Free and Space Bunny Free are all described as zero-retention and not used for training. This is the cleanest tier on the page.

• Not used for training, but retained 30 days — Beam Free sits here, alongside LL Alpha Free. OpenCode's wording is "Prompts and other inputs are not used for training. Data is retained for 30 days for safety reasons," and it repeats the same wording in the free-models note. A separate group, the N​emotron free endpoints, carries a harder warning still: "Trial use only — do not submit personal or confidential data."

That places Beam Free in the middle of the offer list, and the reason it matters more here than it would for a chat product is the harness. A coding agent does not send a question; it sends the files it has read, the diffs it has generated, and the surrounding context it decided was relevant — often several hundred thousand tokens of it, which is the entire point of a 1M window. Every one of those bytes lands in the 30-day bucket. If your repository is the thing you would not paste into a public issue tracker, the free tier's retention line is the term worth negotiating, and the negotiation available to you is: run Beam 1.0 from the Apache 2.0 weights on your own hardware, where no third party retains anything.

The block list is longer than the offer is useful

Beam Free is also the most geographically restricted model on OpenCode's free list, and the restriction is worth reading in full before anyone plans around it.

• Unavailable in — Angola, Belarus, Cambodia, China including Hong Kong and Macau, Cuba, the Democratic Republic of the Congo, Equatorial Guinea, Iran, Iraq, Laos, Libya, Nigeria, North Korea, Russia, Syria, Turkmenistan, Uzbekistan, Venezuela, and the Crimea, Donetsk and Luhansk regions of Ukraine. That is eighteen countries plus three named regions.

• What the block actually attaches to — OpenCode attributes it to the provider behind Beam Free, not to the model. That distinction is doing real work: Beam 1.0's weights are Apache 2.0 and downloadable without a regional gate, so a team in a blocked country can still run the checkpoint. What it cannot do is use this particular free endpoint, and the fallback is self-hosting or Reflection's own API rather than a different free tier.

• US and EU hosting — OpenCode states that its models are hosted in the US and EU, which is the other half of the same question for anyone with data-residency obligations. A US-and-EU-hosted endpoint with a 30-day retention window and an eighteen-country block is a specific compliance shape, and it is checkable today rather than after procurement.

What the free window is worth against a real price

Free is only informative against something, and for Beam 1.0 that something now exists. Because the weights are Apache 2.0, third-party providers serve the model commercially, and Artificial Analysis — which benchmarks it independently rather than repeating Reflection's own numbers — records a median of $0.70 per million input tokens and $2.10 per million output tokens across the providers it tracks, a blended $0.43 per million at its standard 7:2:1 cache-hit mix, and $0.57 to run one Intelligence Index task. So the offer is a genuine discount against a published anchor, not the only number in existence. What no one has published is a first-party rate card: nothing on the record says what Reflection itself will charge for Beam 1.0 once OpenCode's window closes.

That reframes the risk in a way worth being precise about. Zero against $0.70 and $2.10 is a real saving, and if the free route ends there is a priced fallback already in place: the third-party endpoints serving these weights at around that median, or the weights on your own hardware. What is genuinely unmodellable is the other direction — Reflection publishing first-party rates that undercut or replace the third-party market, which changes the comparison for everyone serving the model at once. And the workload widens the range either way: a model that always reasons, defaults to its highest reasoning effort and writes long traces is a different line item at $2.10 per million output tokens than at the $15.00 some flagships charge, so the number to plan against is the one you would pay after the window, not the zero you pay now.

Beam 1.0 is not one of the models we serve, and nothing here is an availability claim from us; the ways to call it today are OpenCode's free endpoint, Reflection's own API, or the Apache 2.0 weights on your own hardware. What we can do is take the second half of that risk off the table. OrcaRouter is one OpenAI-compatible key in front of 200-plus models with the provider's list price passed through at 0% markup, which means a vendor price change is live on the same day rather than on a repricing cycle, and automatic failover means the day a free tier ends or an endpoint has a bad afternoon is a routing change rather than a migration. If Beam is your primary and you want a priced model behind it, these are the three on our catalogue that share its shape — long context, tool calling, structured output:

• GLM-5.2 — GLM-5.2 at $1.40 and $4.40 per million tokens with a 1,000,000-token window, the closest thing on our list to a like-for-like long-context agent model.

• Kimi K3 — 1,048,576-token context at $3.00 per million input tokens, $0.30 cached input and $15.00 output, the flagship tier if the pilot needs to beat Beam rather than match it.

• Qwen3.8-Max — a 1,000,000-token window at $2.00 per million input and $6.00 output, with open weights and a published rate card in both directions.

What a free 1M-token coding window is actually good for

The specs Beam Free exposes through OpenCode are the ones that change how a coding agent behaves, and it is worth being precise about which of them you get through this route and which you get from the model.

• Context — 1,048,576 tokens, the full window. There is no reduced-context free variant here, which is the single most useful thing about this offer: the free tier is the whole model, not a trimmed one.

• Output ceiling — 131,072 tokens on the model, with OpenCode's own request shaping layered on top. If your agent emits large diffs or long reasoning traces, check the effective ceiling on the endpoint rather than assuming the model card's number survives the gateway intact.

• Reasoning — Beam 1.0 reasons by default rather than on request, with effort levels running from max down to low, and max as the default. That is what makes a free window more valuable than usual: an always-on reasoning model burns thinking tokens on every call, and free thinking tokens are the reason to spend the effort budget rather than the reason to ration it. It is also the setting most likely to make the endpoint feel slow, so latency is what you pay instead of money.

• Independent measurement — Artificial Analysis runs Beam 1.0 at an Intelligence Index of 30, which it places 19th of the 120 models it ranks, and at 227.9 output tokens per second, second of the 77 models in its speed ranking — the second figure is a median across the providers serving the weights, not a Reflection number. All of that is third-party. The architectural numbers are not: 501.3B total parameters and 23.4B active per token are vendor-reported and unreproduced by us.

The three things to settle before the window closes

The first is whether you are allowed to use it. Run your team's locations against the eighteen-country block and your obligations against the 30-day retention line before you point an agent at it, because both are published and neither is negotiable on this route. The zero-retention free models on the same page are the alternative if retention is the blocker, and the Apache 2.0 weights are the alternative if the geography is.

The second is what your fallback costs. Beam 1.0 does have a market price — the third-party median Artificial Analysis records — but no first-party one, so the useful preparation is to know today what the nearest priced model costs per million tokens on the same workload. Run that comparison through one key without a second contract and an unannounced ending becomes a config change rather than a scramble.

The third is watching for the two dates that will settle this. One is OpenCode publishing an end date for Beam Free, which would convert an open-ended promotion into a countdown and is the single most useful thing it could add to that table. The other is Reflection publishing a first-party rate card for Beam 1.0, which would put a ceiling on the third-party median and make the keep-or-drop decision an arithmetic question rather than a guess. Until one of those lands, the honest summary is that Beam 1.0 is free inside OpenCode, unlimited by any published measure, restricted by geography, retained for thirty days, and available on terms nobody has committed to beyond "a limited time".

The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.