Morph launches $200 Ultra plan with unlimited Kimi and DeepSeek

• Morph Ultra offers Kimi and DeepSeek for $200/month, initially limited to 1,000 subscriptions, according to Morph's announcement.
• Cached input is unlimited, as skeptrune emphasized, but the plan documentation sets per-minute limits on both uncached input and output. Shared-capacity throttling also applies.
• The advertised $42,000 of tokens for $200 depends on a 99% cache-hit scenario and Morph's own list prices, as shown in skeptrune's infographic.

The published limits include output tokens, contradicting the launch claim that only uncached prefill would be throttled. Morph also publishes hourly serving telemetry and documents a thinking-replay requirement that can trip up custom coding harnesses.

Morph's coverage table lists three included API IDs:

• Kimi K3.1 Pro: morph-kimik31, with the Pro designation confirmed in skeptrune's correction.
• DeepSeek V4.1 Flash: morph-dsv41flash.
• DeepSeek V4.1 Flash 1017: morph-dsv41flash-1017.

Included models have no monthly token cap. Every other model, explicitly including morph-dsv41flash-abliterated, retains per-token billing at list price; Ultra includes no credits for that usage, according to the same documentation.

Morph added storage to its clusters to support cache-heavy workloads, skeptrune said when pitching the plan to developers doing long-context coding work.

The infographic gives six claimed token-value multiples relative to the $200 subscription, using Morph's list prices:

• 70% cache hits: 70×.
• 80% cache hits: 74×.
• 90% cache hits: 88×.
• 95% cache hits: 102×.
• 97% cache hits: 117×.
• 99% cache hits: 210×, or $42,000 in list-priced tokens.

Cached input sits outside the per-minute token budgets, the feature skeptrune defended in a reply about unlimited caching. The remaining uncached input and generated output still consume their respective rate budgets.

Morph's K3.1 pricing guide models a 30-turn agent session with approximately 94% cached input and 60,000 output tokens. Its comparisons expose how much the baseline changes the apparent saving:

• All cache misses → modeled cache reuse: $8.00 → $2.04 per session at Morph's default rates, −$5.96 per session.
• Morph default → Morph standby: $2.04 → $1.02 for the cached session, −50%.
• Moonshot direct → Morph default: $1.48 → $2.04 for the cached session, +$0.56 per session.

These are calculated workloads, not measured Ultra usage. The guide prices K3.1 Pro on Morph at $2.50 per million uncached input tokens, $0.29 cached, and $14 output; its Moonshot comparison uses $2, $0.20, and $10 respectively.

Standby is a separate metered tier that runs on spare capacity and can return a fast 429 when busy. The $42,000 marketing comparison uses Morph's list-price baseline rather than Moonshot's cheaper direct pricing.

Morph published a status page built with Rune from the OpenCode team after users questioned whether unlimited usage would slow the service.

The dashboard snapshot, updated November 10 at 00:28 UTC, reports these medians for the hour beginning November 9 at 23:00 UTC:

| Model | Output speed | Time to first token |
| --- | ---: | ---: |
| Kimi K3.1 | 129 tokens/s | 1.67 seconds |
| DeepSeek V4.1 Flash | 156 tokens/s | 0.74 seconds |

The page describes measurements across all traffic on each model, shared with Ultra. Output speed is the reciprocal of the median gap between tokens; both metrics are measured on Morph's servers and exclude client-network latency.

Skeptrune said the only rate limits would apply to uncached prefill tokens. The published documentation specifies separate output limits too:

| Model | Uncached input tokens/minute | Output tokens/minute |
| --- | ---: | ---: |
| Kimi K3.1 | 72,870 | 6,162 |
| DeepSeek V4.1 Flash, both versions | 2,103,237 | 76,559 |

The documented accounting rules are unusually specific:

• Cached input counts toward neither limit.
• A request larger than a limit still runs; subsequent requests wait for the budget to recover.
• Output is counted when a response finishes.
• The two DeepSeek variants share one set of limits.
• A 429 identifies the reached limit, and Retry-After specifies the wait.

Kimi's output budget is the consequential fine print for parallel agents: 6,162 tokens/minute averages 102.7 tokens/s, below the dashboard's 129 tokens/s streaming median. At that measured speed, about 47.8 seconds of generation produces one minute's output allowance.

Skeptrune said Morph had capacity for 1,000 users without rate limiting, while Morph later pointed a questioner to the rate limits. The documentation reserves no capacity for Ultra and permits capacity-based 429 responses even apart from the token budgets; reserved capacity with an availability SLA belongs to dedicated endpoints.

Morph's integration guide serves OpenAI-compatible requests and Anthropic's Messages format. Its Claude Code example maps Kimi to the main model and DeepSeek to the small-model slot:

The same guide documents several harness-level details:

• Codex transport: /v1/responses, with wire_api = "responses" in its provider configuration.
• Thinking replay: K3.1 Pro requires the complete assistant message on subsequent turns, including reasoning_content and tool_calls. Keeping only content discards earlier reasoning.
• Worker affinity: run_id and x-session-id keep related requests on the worker holding their cached prefix.
• Region placement: prompt_cache_key, already sent by Codex and OpenCode, steers region placement.

Ultra is US-hosted, according to skeptrune, and uses Morph's standard data-retention terms, as another reply confirmed.

The subscription terms distinguish account coverage and plan transitions:

• Personal subscriptions cover personal API keys.
• Organization subscriptions are billed per member, cover organization keys, and pool members under one set of rate limits.
• Existing subscriptions are replaced. The unused portion reduces the first charge, but remaining subscription credits do not carry over.
• After cancellation takes effect, usage of the included models returns to per-token billing.