feat: add hyper charm provider

[oskarkocol]

@rekram1-node before merging, make sure you add a repo secret HYPER_API_KEY. You can get it here: https://hyper.charm.land/. I had no issues making one - no credit card was needed.

Context

Our friends at charm land shipped a gateway.

I'm adding them as a provider.
And also closing: #2747
Their prices are accessible via API therefore adding automation:

curl -X GET https://hyper.charm.land/v1/models \
     -H "Authorization: Bearer sk-hyper-XYZ" \
     -H "Content-Type: application/json"

Tests

Local

bun test packages/core/test/sync.test.ts --filter "Hyper"

Good. Screemshot proof:

bun models:sync hyper --dry-run

Good also.

bun models:sync hyper

Good. Great.

Github Action test on a fork

Created a simple action just to test the functionality. This branch on my fork. https://github.com/oskarkocol/models.dev-contrib/tree/test%2Fhyper-sync-action

https://github.com/oskarkocol/models.dev-contrib/actions/runs/28957929117/job/85921707951

Looks good.

[oskarkocol]

Logo from the Press kit: https://charm.land/presskit

[rekram1-node]

can u add a sync automation for it too?

[rekram1-node]

[github-actions]

I have confirmed the key findings. Here is my review.

Action items

• [high] [violation] packages/core/src/sync/providers/hyper.ts (buildCost) - Check: Sync must preserve hand-authored cost fields the provider API is not authoritative for. Why: buildCost returns only input, output, cache_read, and reasoning, dropping cache_write (and input_audio, output_audio, tiers) from existing TOMLs whenever the API exposes input/output pricing. The sync run in this PR deleted cache_write from minimax-m2.7 (0.375), qwen3.6-flash (0.1), qwen3.6-max (0.2), and qwen3.6-plus (0.2) — the very values this PR introduced. sync.md requires preserving hand-authored fields the API does not expose, as Venice/OpenRouter do. Action: Preserve existing.cost.cache_write (plus input_audio, output_audio, tiers) in buildCost and restore the deleted cache_write values.
• [medium] [possible mistake] providers/hyper/models/gemma-4-26b-a4b-it.toml and llama-4-maverick-17b-128e-instruct-fp8.toml - Check: Provider modalities must be consistent with attachment. Why: buildHyperModel never emits modalities, so these files now inherit input = ["text", "image"] from their base models while setting attachment = false. The hand-authored files deliberately overrode modalities to text-only; the sync removed that override, producing image input with attachments disabled. Action: Verify whether Hyper serves image input for these models; if not, emit a text-only modalities override (or otherwise reconcile attachment).
• [medium] [violation] .github/workflows/pricing-updater.yml:35 (and the update job checkout) - Check: Third-party actions in new automation must be pinned to full commit SHAs. Why: actions/checkout@v4 is referenced by tag, not SHA, which contradicts sync.md ("Actions are pinned by commit SHA") and the PR's own test-hyper-sync.yml (which pins actions/checkout@34e1148…). Action: Pin both actions/checkout uses to a full commit SHA.
• [medium] [possible mistake] packages/core/src/sync/providers/hyper.ts (reasoningOptions) and the synced toggle models - Check: type = "toggle" requires evidence of the exact request field and on/off values. Why: The sync defaults every reasoning model without effort levels to toggle (~14 models), but neither the PR body nor the TOMLs document the actual toggle request syntax. The audit-reasoning-options evidence standard requires the exact field; supports_reasoning alone proves reasoning exists, not a caller-controlled toggle. Action: Cite and document the exact toggle request field/values, or use reasoning_options = [] for models where no toggle control is verified.
• [low] [possible mistake] providers/hyper/models/minimax-m2.7.toml ([cost]) - Check: Verify the cost_per_1m_in_cached → cache_read mapping. Why: cache_read changed from 0.06 to 0.375, which was the previous cache_write value, while cache_write was deleted — suggesting cost_per_1m_in_cached may represent the cache-write price rather than read. The sync also sources pricing from /v1/provider, an endpoint not cited in the PR body (which only cites /v1/models). Action: Confirm against the Hyper API which field is cache-read vs cache-write, correct the mapping if needed, and cite /v1/provider if it is the pricing source.
• [low] [possible mistake] providers/hyper/models/qwen3.6-max.toml - Check: Provider-specific name overrides must survive sync. Why: The hand-authored name = "Qwen3.6-Max" was removed by the sync, so the model now displays as the base's "Qwen3.6 Max Preview" despite being filed under qwen3.6-max. buildHyperModel does not emit or preserve name. Action: Preserve the existing name override or confirm "Qwen3.6 Max Preview" is the intended display name.
• [low] [possible mistake] .github/workflows/test-hyper-sync.yml - Check: Do not merge fork-only test workflows into upstream. Why: The workflow is guarded to run only on forks and its own header says "Do not merge into upstream PRs"; it is a test artifact for validating the sync on a fork. Action: Remove this workflow before merging.

[oskarkocol]

• Fixed cost sync to keep API pricing and preserve fields Hyper doesn't expose
• Text-only modalities when attachments are disabled; removed unverified toggle reasoning options.
• Re-synced all 20 Hyper models; validation passes
• Removed fork-only test workflow - that was a bad merge conflict resolve on my part

[github-actions]

Action items

• [high] [possible mistake] packages/core/src/sync/providers/hyper.ts (buildCost, fromProviderFields) - Check: Whether Hyper's cost_per_1m_in_cached and cost_per_1m_out_cached API fields map to cache_read/cache_write correctly. Why: cost_per_1m_in_cached is mapped to cache_read and cost_per_1m_out_cached to cache_write, but the synced values are reversed from the standard 10%/125% (read/write) ratio documented in AGENTS.md. For qwen3.6-flash.toml, cache_read = 1.25 (125% of input 1) and cache_write = 0.1 (10%) — the expensive value landed in cache_read and the cheap value in cache_write. The same inversion appears in qwen3.6-max, qwen3.7-max, and qwen3.7-plus. Strongest evidence: the hand-authored minimax-m2.7.toml initially had the correct ratio (cache_read = 0.06 at 20%, cache_write = 0.375 at 125%), but after the first sync its cache_read became 0.375 — the old cache_write value — confirming cost_per_1m_in_cached carries the write price. Additionally, fromProviderFields (the swapped path) takes precedence over the correctly-mapped fromPricingObject path (input_cache_read→cache_read, input_cache_write→cache_write), so the inversion applies whenever the cost_per_1m_* fields are populated. Action: Verify the Hyper API's field semantics against provider documentation. If cost_per_1m_in_cached is the cache-write price, swap the mapping so it targets cache_write and cost_per_1m_out_cached targets cache_read, then re-sync.

• [medium] [possible mistake] providers/hyper/provider.toml:8-9 - Check: Whether Hyper exposes a per-request reasoning toggle for models without effort levels, and whether the provider.toml comment contradicts the sync's reasoning_options output. Why: The comment states "models without effort support expose reasoning as a toggle only," but the sync returns reasoning_options = [] for 14 such models, with a code comment stating "Hyper advertises reasoning but documents no toggle or effort control for these models." If Hyper does expose a caller-facing reasoning toggle, those 14 models need [{ type = "toggle" }] reasoning_options; if not, the provider.toml comment is misleading because "toggle" in this repo's schema specifically means a caller-facing on/off control. Action: Verify whether Hyper's API accepts a per-request reasoning toggle parameter for models where supports_reasoning_effort is false. If yes, update reasoningOptions() to emit { type: "toggle" } and re-sync; if no, reword the provider.toml comment to clarify that supports_reasoning is an advertisement flag, not a caller-facing control.

[oskarkocol]

sure, let's clarify the comment

[oskarkocol]

PR reviewer ded. Gotta rerun the action

[andreynering]

Hello! 👋

Thank you both for the interest in listing Hyper on models.dev.

I would just like to let you know that we're planning some small changes to our /models endpoint for soon:

• We plan to make costs available for public requests (so you won't need an API key for that anymore).
• We'll probably rename some fields for clarity, in particular the costs because cache_in and cache_out are not so clear, and should be something like cache_creation and cache_hit instead.

So it's probably a good idea to hold this for a bit, and I'll let you know here once we do these changes.

By the way, good timing to let us know if you miss anything on the endpoint.

[oskarkocol]

Thanks for the heads-up @andreynering.

I'm going to close this PR. We can revisit once Hyper's new /models endpoint ships and is stable. No sense in merging this if the pricing fields (cache_in/cache_out`) and auth requirements are changing in the near term

[andreynering]

IMO it’s worth keeping it open as a draft, but it’s your choice.

[oskarkocol]

Prefer closing than letting it go stale. Drop a comment here when the new endpoint is live and we can easily spin up a fresh one!

[andreynering]

@oskarkocol Updated endpoint is live: https://hyper.charm.land/v1/models