Deepseek API price increase: Alternatives for your agents

DeepSeek has now published the replacement prices for its API. From August 16, 2026, DeepSeek V4 Flash will use peak and off-peak billing. The cheapest new rate is higher than the old rate, and the peak rate is substantially higher. The decision is no longer whether to prepare for an unspecified increase. It is whether to schedule work around DeepSeek's off-peak hours, move suitable jobs to another hosted model, or use more than one provider.

Three alternatives deserve particular attention. Muse Spark 1.2 Contributor is the cheapest on direct token cost, but only if you can permit Meta to use prompts and completions to improve future models. Tencent Hy3 undercuts DeepSeek's new off-peak Flash price without that specific contributor-tier tradeoff and provides the core interfaces an agent needs. GPT-5.6 Luna costs more per token, but its hosted tool surface can reduce the amount of agent infrastructure you operate yourself.

The right metric is still cost per verified result, not cost per token. Keep tools, authorization, state, approvals, and evaluations independent of the model so that routing remains reversible.

DeepSeek's new peak and off-peak prices

DeepSeek says the following rates take effect at 16:00 UTC on August 16, 2026. Peak hours are 01:00-04:00 UTC and 06:00-10:00 UTC. All other hours are off-peak.

DeepSeek API model and period	Cached input per 1M tokens	Uncached input per 1M tokens	Output per 1M tokens
V4 Flash, old price	$0.0028	$0.14	$0.28
V4 Flash, new off-peak	$0.007	$0.22	$0.66
V4 Flash, new peak	$0.014	$0.44	$1.32
V4 Pro, old price	$0.003625	$0.435	$0.87
V4 Pro, new off-peak	$0.022	$0.66	$1.98
V4 Pro, new peak	$0.044	$1.32	$3.96

The off-peak rate is exactly half the peak rate, but it is not a discount from the old price. For V4 Flash, off-peak uncached input rises by about 57%, cached input by 150%, and output by about 136%. At peak, uncached input and output are roughly 3.1 and 4.7 times their old prices. DeepSeek's Models & Pricing page publishes the rates, time windows, and effective date.

This creates an architectural choice that did not exist under flat pricing. Interactive work must usually run when the user asks. Batch extraction, evaluation replays, indexing, summarization, and background enrichment can often be queued for off-peak execution. A router should record which price window applied to every request rather than assigning one blended DeepSeek rate after the fact.

What the increase means for a real agent job

Suppose one completed job consumes:

• 100,000 uncached input tokens;
• 20,000 cached input tokens; and
• 10,000 output tokens.

The old DeepSeek V4 Flash cost was:

(0.10 × $0.14) + (0.02 × $0.0028) + (0.01 × $0.28) = $0.016856

The same job is now $0.02874 off-peak and $0.05748 at peak. At 10,000 jobs per month, direct model charges rise from $168.56 to between $287.40 and $574.80, depending on when every request runs. If traffic were spread evenly through the day, the time-weighted model cost would be about $371.23 per month, before retries, tools, infrastructure, or review.

Here is the same workload using each provider's published cache rate. This is an estimate, not a guaranteed invoice: providers do not necessarily identify or retain the same cacheable prefixes.

Candidate	Published input / cached input / output price per 1M tokens	Cost per illustrative job	Cost for 10,000 jobs
Muse Spark 1.2 Contributor	$0.10 / $0.002 / $0.20	$0.01204	$120.40
Tencent Hy3 on international TokenHub	$0.132 / $0.033 / $0.528	$0.01914	$191.40
DeepSeek V4 Flash, off-peak	$0.22 / $0.007 / $0.66	$0.02874	$287.40
GPT-5.6 Luna	$0.20 / $0.02 / $1.20	$0.03240	$324.00
DeepSeek V4 Flash, peak	$0.44 / $0.014 / $1.32	$0.05748	$574.80

The table is useful for deciding what to test. It is not a model ranking. A cheap model that needs more turns, produces invalid tool arguments, or requires human repair can cost more per completed action. Reasoning tokens, hosted tools, web search, retries, and long-context surcharges also need to be included where they apply.

Muse Spark 1.2 Contributor: the cheapest option has a data-use condition

Meta lists two commercial routes for Muse Spark 1.2. The standard model costs $1.25 for input, $0.15 for cached input, and $4.25 for output per million tokens. The Contributor model costs $0.10, $0.002, and $0.20 respectively. The low rate is offered in exchange for permission to use submitted prompts and completions to train future Meta models. Meta's Muse Code announcement describes the Contributor tier and its pricing.

That condition should determine eligibility before quality or price. Contributor is a plausible option for public open-source code, synthetic evaluations, disposable prototypes, and other material that the organization has deliberately approved for contribution. It is usually not an option for customer records, private repositories, personal data, secrets, unreleased product plans, regulated data, or content covered by contractual restrictions. Redaction helps, but it is not a substitute for an explicit data classification and approval policy.

Do not send a job to Contributor merely because a model router predicts that it will be cheaper. Put the data-use decision in the application policy:

1. Classify the request and every retrieved item before model invocation.
2. Allow Contributor only for approved data classes.
3. Prevent tools from adding restricted data later in the run.
4. Record the selected tier and policy decision in the audit log.
5. Fail closed to a non-contributor endpoint when classification is uncertain.

Muse Spark is designed for agentic and multimodal work and is available through Meta's OpenAI-compatible Model API. Meta currently describes the API as a public preview for US developers, so verify account and regional availability before treating it as a production route. The Contributor price should be evaluated as a separate product contract, not as a universally cheaper version of DeepSeek. If the data condition is unacceptable, the relevant comparison is standard Muse Spark at $1.25/$0.15/$4.25, which is far more expensive than DeepSeek Flash.

Tencent Hy3: the closest conventional price alternative

Tencent's production hy3 model is the strongest direct price comparison in this group. Tencent Cloud's international TokenHub lists $0.132 input, $0.033 cached input, and $0.528 output per million tokens. Its mainland China price is ¥1, ¥0.25, and ¥4. Tencent Cloud's international pricing page and China pricing page publish the regional rates.

Hy3 has a 256K context window, up to 192K input and 128K output, retained reasoning, structured output, function calling, and caching. Tencent documents support for OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages on TokenHub. The model list and protocol overview define that interface.

Those capabilities make Hy3 more than a cheap classifier. It is a candidate for multi-step tool use, repository work, document review, and read-and-propose agents. It also has published open weights under Apache 2.0, so a successful API evaluation can inform a later self-hosting investigation without making self-hosting the first step.

There are two constraints to test explicitly. First, 256K is materially smaller than DeepSeek's 1M context, so an agent that depends on sending an entire repository or long history may need retrieval, compaction, or state summarization. Second, regional availability, account setup, data residency, support, and contract terms can matter more than the token price. Test the endpoint and commercial region you would actually use; do not combine a mainland price, an international latency measurement, and a third-party model benchmark into one supposed result.

Hy3 is part of a broader Chinese-model price category rather than an isolated exception. Qwen, MiniMax, Kimi, GLM, and DeepSeek models are all available through first-party or aggregation APIs, often with different context tiers and cache rules. Use Hy3 as the first conventional price test for this workload, then add other Chinese models only when they have a credible fit for the same tool and data contract. A long list of cheap model names is less useful than three candidates evaluated on identical traces.

GPT-5.6 Luna: pay more for the surrounding agent platform

GPT-5.6 Luna remains priced at $0.20 input, $0.02 cached input, and $1.20 output per million tokens. OpenAI lists a 1.05M-token context, 128K maximum output, function calling, structured outputs, image input, and reasoning support. Through the Responses API it can use web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. OpenAI's GPT-5.6 Luna model page is the source for the rates and supported surface.

Luna is slightly more expensive than DeepSeek's new off-peak rate for the illustrative cached workload, but cheaper than DeepSeek at peak. More importantly, its value may come from the platform rather than text generation alone. If hosted file search, shell execution, computer use, MCP, or tool search replaces infrastructure you would otherwise build and operate, Luna can win at the workflow level despite a higher token bill.

Include tool charges and the tokens returned by tools in the ledger. Long prompts above 272K also use higher rates under OpenAI's published pricing rules. For a fair test, compare the complete run: request, reasoning, tool calls, retries, final state, and review—not merely 120,000 text tokens passed through each model.

The migration detail is API shape. DeepSeek, Tencent TokenHub, Meta Model API, and OpenAI all provide familiar interfaces, but their reasoning items, cache controls, tool events, and continuation semantics differ. Keep an internal provider-neutral event model—a user message, model output, tool call, tool result, approval decision, usage record, and final product state—and translate at the edge.

Self-hosting is a separate economic decision

DeepSeek V4 Flash and Tencent Hy3 both have published weights. That does not turn either one into free inference. The hosted token bill is replaced by accelerators, idle capacity, quantization decisions, serving software, deployment, monitoring, security, model updates, and on-call work.

Self-hosting becomes worth investigating when utilization is high and predictable, the serving environment must be controlled, or the team already operates GPU inference. It is usually a distraction when the immediate problem can be solved by moving batch jobs off-peak or routing an approved workload to another hosted endpoint. Benchmark the exact weights, precision, hardware, concurrency, and context length before comparing self-hosted cost with an API price.

Preserve the workflow contract during migration

The portable unit is the workflow, not the prompt. Define it as:

1. A tenant-scoped request and the minimum context required for the current step.
2. A versioned set of read, propose, and write tools.
3. Server-side validation and authorization for every tool argument.
4. An approval state for consequential actions.
5. A durable record of model events, tool calls, price window, and resulting state.
6. A finish condition that can be checked without trusting the model's explanation.

Put provider-specific translation around message formats, tool schemas, structured-output parsing, reasoning settings, caching, retries, cancellation, and usage accounting. Keep permissions in the service that owns the action. A model may select a tool; it does not grant itself authority to execute it.

Run an evaluation that can change the decision

Replay privacy-safe traces from one workflow. A useful first set has 40 cases:

• 20 ordinary requests;
• 5 incomplete-context cases;
• 5 ambiguous record matches;
• 5 denied or cross-tenant access attempts; and
• 5 tool failures, retries, duplicate requests, or rejected approvals.

Run every candidate with the same retrieval results, tools, approval policy, maximum turns, and post-action verifier. Record verified completion, unauthorized calls, valid tool arguments, retries, total turns, token categories, provider and tool charges, latency, and human edits. For DeepSeek, run or replay both price windows. For Muse Contributor, use only traces approved for contribution. For Hy3, include cases that expose the effect of its smaller context window.

Set the release rule before looking at results. For example: no unauthorized writes, at least 95% verified completion on ordinary cases, no more than two retries per job, and a lower total cost per verified result than the current route. Start in shadow or draft-only mode, then promote a small percentage of traffic while retaining a fallback.

Where Ginger Labs fits

At Ginger Labs, we build embedded AI agents and copilots for SaaS products. The agent can live in a side panel, inline surface, or modal and reason over the customer's schemas, records, stages, and rules to progress defined multi-step work. The SDK includes retrieval, evaluations, self-learning loops, and observability.

The same workflow can be evaluated on DeepSeek, Muse Spark, Tencent Hy3, GPT-5.6 Luna, or a self-hosted endpoint. The customer retains its API, data model, permissions, tenant boundaries, action policy, user experience, and definition of a correct result. Ginger Labs supplies the embedded agent layer around that contract while the model remains a swappable, measurable component.

Recommendation

Do not perform a blind one-model migration. Use a routing policy:

1. Keep latency-sensitive DeepSeek work on its current route while measuring the new peak and off-peak cost separately.
2. Move safe batch work to DeepSeek off-peak hours when the operational delay is acceptable.
3. Test Tencent Hy3 first as the closest conventional hosted price alternative.
4. Test Muse Spark 1.2 Contributor only on data explicitly approved for use in future model training; otherwise compare against standard Muse pricing.
5. Test GPT-5.6 Luna when its hosted tools can replace application infrastructure or when jobs must run during DeepSeek peak hours.
6. Consider self-hosting only after measured, predictable utilization justifies an inference operation.

DeepSeek's increase does not produce one universal replacement. It makes time, data rights, regional deployment, tool infrastructure, and completion quality part of the price calculation. The best route is the one that meets the workflow's completion and authorization thresholds at the lowest cost per verified result.

For the product architecture behind this evaluation, see What Matters Most When Building AI Agents for Business Workflows.

Sources

• Models & Pricing and Tool Calls, DeepSeek. Accessed August 16, 2026.
• Build with Muse Code and Meta Model API overview, Meta. Accessed August 16, 2026.
• International model pricing, China model pricing, model list, protocol overview, and Hy3 model card, Tencent. Accessed August 16, 2026.
• GPT-5.6 Luna model page, OpenAI. Accessed August 16, 2026.
• DeepSeek V4 Flash model card, DeepSeek-AI. Accessed August 16, 2026.
• Ginger Labs, product overview. Accessed August 16, 2026.